8 5. Detailed list of parameters
12 9. CPU idleness profiling
13 10. Verification and triggers
17 1.0 Overview and history
18 ------------------------
19 fio was originally written to save me the hassle of writing special test
20 case programs when I wanted to test a specific workload, either for
21 performance reasons or to find/reproduce a bug. The process of writing
22 such a test app can be tiresome, especially if you have to do it often.
23 Hence I needed a tool that would be able to simulate a given io workload
24 without resorting to writing a tailored test case again and again.
26 A test work load is difficult to define, though. There can be any number
27 of processes or threads involved, and they can each be using their own
28 way of generating io. You could have someone dirtying large amounts of
29 memory in an memory mapped file, or maybe several threads issuing
30 reads using asynchronous io. fio needed to be flexible enough to
31 simulate both of these cases, and many more.
35 The first step in getting fio to simulate a desired io workload, is
36 writing a job file describing that specific setup. A job file may contain
37 any number of threads and/or files - the typical contents of the job file
38 is a global section defining shared parameters, and one or more job
39 sections describing the jobs involved. When run, fio parses this file
40 and sets everything up as described. If we break down a job from top to
41 bottom, it contains the following basic parameters:
43 IO type Defines the io pattern issued to the file(s).
44 We may only be reading sequentially from this
45 file(s), or we may be writing randomly. Or even
46 mixing reads and writes, sequentially or randomly.
48 Block size In how large chunks are we issuing io? This may be
49 a single value, or it may describe a range of
52 IO size How much data are we going to be reading/writing.
54 IO engine How do we issue io? We could be memory mapping the
55 file, we could be using regular read/write, we
56 could be using splice, async io, or even SG
59 IO depth If the io engine is async, how large a queuing
60 depth do we want to maintain?
62 IO type Should we be doing buffered io, or direct/raw io?
64 Num files How many files are we spreading the workload over.
66 Num threads How many threads or processes should we spread
69 The above are the basic parameters defined for a workload, in addition
70 there's a multitude of parameters that modify other aspects of how this
76 See the README file for command line parameters, there are only a few
79 Running fio is normally the easiest part - you just give it the job file
80 (or job files) as parameters:
84 and it will start doing what the job_file tells it to do. You can give
85 more than one job file on the command line, fio will serialize the running
86 of those files. Internally that is the same as using the 'stonewall'
87 parameter described in the parameter section.
89 If the job file contains only one job, you may as well just give the
90 parameters on the command line. The command line parameters are identical
91 to the job parameters, with a few extra that control global parameters
92 (see README). For example, for the job file parameter iodepth=2, the
93 mirror command line option would be --iodepth 2 or --iodepth=2. You can
94 also use the command line for giving more than one job entry. For each
95 --name option that fio sees, it will start a new job with that name.
96 Command line entries following a --name entry will apply to that job,
97 until there are no more entries or a new --name entry is seen. This is
98 similar to the job file options, where each option applies to the current
99 job until a new [] job entry is seen.
101 fio does not need to run as root, except if the files or devices specified
102 in the job section requires that. Some other options may also be restricted,
103 such as memory locking, io scheduler switching, and decreasing the nice value.
108 As previously described, fio accepts one or more job files describing
109 what it is supposed to do. The job file format is the classic ini file,
110 where the names enclosed in [] brackets define the job name. You are free
111 to use any ascii name you want, except 'global' which has special meaning.
112 A global section sets defaults for the jobs described in that file. A job
113 may override a global section parameter, and a job file may even have
114 several global sections if so desired. A job is only affected by a global
115 section residing above it. If the first character in a line is a ';' or a
116 '#', the entire line is discarded as a comment.
118 So let's look at a really simple job file that defines two processes, each
119 randomly reading from a 128MB file.
121 ; -- start job file --
132 As you can see, the job file sections themselves are empty as all the
133 described parameters are shared. As no filename= option is given, fio
134 makes up a filename for each of the jobs as it sees fit. On the command
135 line, this job would look as follows:
137 $ fio --name=global --rw=randread --size=128m --name=job1 --name=job2
140 Let's look at an example that has a number of processes writing randomly
143 ; -- start job file --
155 Here we have no global section, as we only have one job defined anyway.
156 We want to use async io here, with a depth of 4 for each file. We also
157 increased the buffer size used to 32KB and define numjobs to 4 to
158 fork 4 identical jobs. The result is 4 processes each randomly writing
159 to their own 64MB file. Instead of using the above job file, you could
160 have given the parameters on the command line. For this case, you would
163 $ fio --name=random-writers --ioengine=libaio --iodepth=4 --rw=randwrite --bs=32k --direct=0 --size=64m --numjobs=4
165 When fio is utilized as a basis of any reasonably large test suite, it might be
166 desirable to share a set of standardized settings across multiple job files.
167 Instead of copy/pasting such settings, any section may pull in an external
168 .fio file with 'include filename' directive, as in the following example:
170 ; -- start job file including.fio --
174 include glob-include.fio
181 include test-include.fio
182 ; -- end job file including.fio --
184 ; -- start job file glob-include.fio --
187 ; -- end job file glob-include.fio --
189 ; -- start job file test-include.fio --
192 ; -- end job file test-include.fio --
194 Settings pulled into a section apply to that section only (except global
195 section). Include directives may be nested in that any included file may
196 contain further include directive(s). Include files may not contain []
200 4.1 Environment variables
201 -------------------------
203 fio also supports environment variable expansion in job files. Any
204 sub-string of the form "${VARNAME}" as part of an option value (in other
205 words, on the right of the `='), will be expanded to the value of the
206 environment variable called VARNAME. If no such environment variable
207 is defined, or VARNAME is the empty string, the empty string will be
210 As an example, let's look at a sample fio invocation and job file:
212 $ SIZE=64m NUMJOBS=4 fio jobfile.fio
214 ; -- start job file --
221 This will expand to the following equivalent job file at runtime:
223 ; -- start job file --
230 fio ships with a few example job files, you can also look there for
233 4.2 Reserved keywords
234 ---------------------
236 Additionally, fio has a set of reserved keywords that will be replaced
237 internally with the appropriate value. Those keywords are:
239 $pagesize The architecture page size of the running system
240 $mb_memory Megabytes of total memory in the system
241 $ncpus Number of online available CPUs
243 These can be used on the command line or in the job file, and will be
244 automatically substituted with the current system values when the job
245 is run. Simple math is also supported on these keywords, so you can
246 perform actions like:
250 and get that properly expanded to 8 times the size of memory in the
254 5.0 Detailed list of parameters
255 -------------------------------
257 This section describes in details each parameter associated with a job.
258 Some parameters take an option of a given type, such as an integer or
259 a string. Anywhere a numeric value is required, an arithmetic expression
260 may be used, provided it is surrounded by parentheses. Supported operators
270 For time values in expressions, units are microseconds by default. This is
271 different than for time values not in expressions (not enclosed in
272 parentheses). The following types are used:
274 str String. This is a sequence of alpha characters.
275 time Integer with possible time suffix. In seconds unless otherwise
276 specified, use eg 10m for 10 minutes. Accepts s/m/h for seconds,
277 minutes, and hours, and accepts 'ms' (or 'msec') for milliseconds,
278 and 'us' (or 'usec') for microseconds.
279 int SI integer. A whole number value, which may contain a suffix
280 describing the base of the number. Accepted suffixes are k/m/g/t/p,
281 meaning kilo, mega, giga, tera, and peta. The suffix is not case
282 sensitive, and you may also include trailing 'b' (eg 'kb' is the same
283 as 'k'). So if you want to specify 4096, you could either write
284 out '4096' or just give 4k. The suffixes signify base 2 values, so
285 1024 is 1k and 1024k is 1m and so on, unless the suffix is explicitly
286 set to a base 10 value using 'kib', 'mib', 'gib', etc. If that is the
287 case, then 1000 is used as the multiplier. This can be handy for
288 disks, since manufacturers generally use base 10 values when listing
289 the capacity of a drive. If the option accepts an upper and lower
290 range, use a colon ':' or minus '-' to separate such values. May also
291 include a prefix to indicate numbers base. If 0x is used, the number
292 is assumed to be hexadecimal. See irange.
293 bool Boolean. Usually parsed as an integer, however only defined for
294 true and false (1 and 0).
295 irange Integer range with suffix. Allows value range to be given, such
296 as 1024-4096. A colon may also be used as the separator, eg
297 1k:4k. If the option allows two sets of ranges, they can be
298 specified with a ',' or '/' delimiter: 1k-4k/8k-32k. Also see
300 float_list A list of floating numbers, separated by a ':' character.
302 With the above in mind, here follows the complete list of fio job
305 name=str ASCII name of the job. This may be used to override the
306 name printed by fio for this job. Otherwise the job
307 name is used. On the command line this parameter has the
308 special purpose of also signaling the start of a new
311 wait_for=str Specifies the name of the already defined job to wait
312 for. Single waitee name only may be specified. If set, the job
313 won't be started until all workers of the waitee job are done.
315 Wait_for operates on the job name basis, so there are a few
316 limitations. First, the waitee must be defined prior to the
317 waiter job (meaning no forward references). Second, if a job
318 is being referenced as a waitee, it must have a unique name
319 (no duplicate waitees).
321 description=str Text description of the job. Doesn't do anything except
322 dump this text description when this job is run. It's
325 directory=str Prefix filenames with this directory. Used to place files
326 in a different location than "./". See the 'filename' option
327 for escaping certain characters.
329 filename=str Fio normally makes up a filename based on the job name,
330 thread number, and file number. If you want to share
331 files between threads in a job or several jobs, specify
332 a filename for each of them to override the default.
333 If the ioengine is file based, you can specify a number of
334 files by separating the names with a ':' colon. So if you
335 wanted a job to open /dev/sda and /dev/sdb as the two working
336 files, you would use filename=/dev/sda:/dev/sdb. On Windows,
337 disk devices are accessed as \\.\PhysicalDrive0 for the first
338 device, \\.\PhysicalDrive1 for the second etc. Note: Windows
339 and FreeBSD prevent write access to areas of the disk
340 containing in-use data (e.g. filesystems).
341 If the wanted filename does need to include a colon, then
342 escape that with a '\' character. For instance, if the filename
343 is "/dev/dsk/foo@3,0:c", then you would use
344 filename="/dev/dsk/foo@3,0\:c". '-' is a reserved name, meaning
345 stdin or stdout. Which of the two depends on the read/write
349 If sharing multiple files between jobs, it is usually necessary
350 to have fio generate the exact names that you want. By default,
351 fio will name a file based on the default file format
352 specification of jobname.jobnumber.filenumber. With this
353 option, that can be customized. Fio will recognize and replace
354 the following keywords in this string:
357 The name of the worker thread or process.
360 The incremental number of the worker thread or
364 The incremental number of the file for that worker
367 To have dependent jobs share a set of files, this option can
368 be set to have fio generate filenames that are shared between
369 the two. For instance, if testfiles.$filenum is specified,
370 file number 4 for any job will be named testfiles.4. The
371 default of $jobname.$jobnum.$filenum will be used if
372 no other format specifier is given.
374 unique_filename=bool To avoid collisions between networked clients, fio
375 defaults to prefixing any generated filenames (with a directory
376 specified) with the source of the client connecting. To disable
377 this behavior, set this option to 0.
379 opendir=str Tell fio to recursively add any file it can find in this
380 directory and down the file system tree.
382 lockfile=str Fio defaults to not locking any files before it does
383 IO to them. If a file or file descriptor is shared, fio
384 can serialize IO to that file to make the end result
385 consistent. This is usual for emulating real workloads that
386 share files. The lock modes are:
388 none No locking. The default.
389 exclusive Only one thread/process may do IO,
390 excluding all others.
391 readwrite Read-write locking on the file. Many
392 readers may access the file at the
393 same time, but writes get exclusive
397 rw=str Type of io pattern. Accepted values are:
399 read Sequential reads
400 write Sequential writes
401 randwrite Random writes
402 randread Random reads
403 rw,readwrite Sequential mixed reads and writes
404 randrw Random mixed reads and writes
405 trimwrite Mixed trims and writes. Blocks will be
406 trimmed first, then written to.
408 For the mixed io types, the default is to split them 50/50.
409 For certain types of io the result may still be skewed a bit,
410 since the speed may be different. It is possible to specify
411 a number of IO's to do before getting a new offset, this is
412 done by appending a ':<nr>' to the end of the string given.
413 For a random read, it would look like 'rw=randread:8' for
414 passing in an offset modifier with a value of 8. If the
415 suffix is used with a sequential IO pattern, then the value
416 specified will be added to the generated offset for each IO.
417 For instance, using rw=write:4k will skip 4k for every
418 write. It turns sequential IO into sequential IO with holes.
419 See the 'rw_sequencer' option.
421 rw_sequencer=str If an offset modifier is given by appending a number to
422 the rw=<str> line, then this option controls how that
423 number modifies the IO offset being generated. Accepted
426 sequential Generate sequential offset
427 identical Generate the same offset
429 'sequential' is only useful for random IO, where fio would
430 normally generate a new random offset for every IO. If you
431 append eg 8 to randread, you would get a new random offset for
432 every 8 IO's. The result would be a seek for only every 8
433 IO's, instead of for every IO. Use rw=randread:8 to specify
434 that. As sequential IO is already sequential, setting
435 'sequential' for that would not result in any differences.
436 'identical' behaves in a similar fashion, except it sends
437 the same offset 8 number of times before generating a new
440 kb_base=int The base unit for a kilobyte. The defacto base is 2^10, 1024.
441 Storage manufacturers like to use 10^3 or 1000 as a base
442 ten unit instead, for obvious reasons. Allow values are
443 1024 or 1000, with 1024 being the default.
445 unified_rw_reporting=bool Fio normally reports statistics on a per
446 data direction basis, meaning that read, write, and trim are
447 accounted and reported separately. If this option is set,
448 the fio will sum the results and report them as "mixed"
451 randrepeat=bool For random IO workloads, seed the generator in a predictable
452 way so that results are repeatable across repetitions.
455 randseed=int Seed the random number generators based on this seed value, to
456 be able to control what sequence of output is being generated.
457 If not set, the random sequence depends on the randrepeat
460 fallocate=str Whether pre-allocation is performed when laying down files.
463 none Do not pre-allocate space
464 posix Pre-allocate via posix_fallocate()
465 keep Pre-allocate via fallocate() with
466 FALLOC_FL_KEEP_SIZE set
467 0 Backward-compatible alias for 'none'
468 1 Backward-compatible alias for 'posix'
470 May not be available on all supported platforms. 'keep' is only
471 available on Linux.If using ZFS on Solaris this must be set to
472 'none' because ZFS doesn't support it. Default: 'posix'.
474 fadvise_hint=bool By default, fio will use fadvise() to advise the kernel
475 on what IO patterns it is likely to issue. Sometimes you
476 want to test specific IO patterns without telling the
477 kernel about it, in which case you can disable this option.
478 If set, fio will use POSIX_FADV_SEQUENTIAL for sequential
479 IO and POSIX_FADV_RANDOM for random IO.
481 fadvise_stream=int Notify the kernel what write stream ID to place these
482 writes under. Only supported on Linux. Note, this option
483 may change going forward.
485 size=int The total size of file io for this job. Fio will run until
486 this many bytes has been transferred, unless runtime is
487 limited by other options (such as 'runtime', for instance,
488 or increased/decreased by 'io_size'). Unless specific nrfiles
489 and filesize options are given, fio will divide this size
490 between the available files specified by the job. If not set,
491 fio will use the full size of the given files or devices.
492 If the files do not exist, size must be given. It is also
493 possible to give size as a percentage between 1 and 100. If
494 size=20% is given, fio will use 20% of the full size of the
495 given files or devices.
498 io_limit=int Normally fio operates within the region set by 'size', which
499 means that the 'size' option sets both the region and size of
500 IO to be performed. Sometimes that is not what you want. With
501 this option, it is possible to define just the amount of IO
502 that fio should do. For instance, if 'size' is set to 20G and
503 'io_size' is set to 5G, fio will perform IO within the first
504 20G but exit when 5G have been done. The opposite is also
505 possible - if 'size' is set to 20G, and 'io_size' is set to
506 40G, then fio will do 40G of IO within the 0..20G region.
508 filesize=int Individual file sizes. May be a range, in which case fio
509 will select sizes for files at random within the given range
510 and limited to 'size' in total (if that is given). If not
511 given, each created file is the same size.
513 file_append=bool Perform IO after the end of the file. Normally fio will
514 operate within the size of a file. If this option is set, then
515 fio will append to the file instead. This has identical
516 behavior to setting offset to the size of a file. This option
517 is ignored on non-regular files.
520 fill_fs=bool Sets size to something really large and waits for ENOSPC (no
521 space left on device) as the terminating condition. Only makes
522 sense with sequential write. For a read workload, the mount
523 point will be filled first then IO started on the result. This
524 option doesn't make sense if operating on a raw device node,
525 since the size of that is already known by the file system.
526 Additionally, writing beyond end-of-device will not return
530 bs=int The block size used for the io units. Defaults to 4k. Values
531 can be given for both read and writes. If a single int is
532 given, it will apply to both. If a second int is specified
533 after a comma, it will apply to writes only. In other words,
534 the format is either bs=read_and_write or bs=read,write,trim.
535 bs=4k,8k will thus use 4k blocks for reads, 8k blocks for
536 writes, and 8k for trims. You can terminate the list with
537 a trailing comma. bs=4k,8k, would use the default value for
538 trims.. If you only wish to set the write size, you
539 can do so by passing an empty read size - bs=,8k will set
540 8k for writes and leave the read default value.
543 ba=int At what boundary to align random IO offsets. Defaults to
544 the same as 'blocksize' the minimum blocksize given.
545 Minimum alignment is typically 512b for using direct IO,
546 though it usually depends on the hardware block size. This
547 option is mutually exclusive with using a random map for
548 files, so it will turn off that option.
550 blocksize_range=irange
551 bsrange=irange Instead of giving a single block size, specify a range
552 and fio will mix the issued io block sizes. The issued
553 io unit will always be a multiple of the minimum value
554 given (also see bs_unaligned). Applies to both reads and
555 writes, however a second range can be given after a comma.
558 bssplit=str Sometimes you want even finer grained control of the
559 block sizes issued, not just an even split between them.
560 This option allows you to weight various block sizes,
561 so that you are able to define a specific amount of
562 block sizes issued. The format for this option is:
564 bssplit=blocksize/percentage:blocksize/percentage
566 for as many block sizes as needed. So if you want to define
567 a workload that has 50% 64k blocks, 10% 4k blocks, and
568 40% 32k blocks, you would write:
570 bssplit=4k/10:64k/50:32k/40
572 Ordering does not matter. If the percentage is left blank,
573 fio will fill in the remaining values evenly. So a bssplit
574 option like this one:
576 bssplit=4k/50:1k/:32k/
578 would have 50% 4k ios, and 25% 1k and 32k ios. The percentages
579 always add up to 100, if bssplit is given a range that adds
580 up to more, it will error out.
582 bssplit also supports giving separate splits to reads and
583 writes. The format is identical to what bs= accepts. You
584 have to separate the read and write parts with a comma. So
585 if you want a workload that has 50% 2k reads and 50% 4k reads,
586 while having 90% 4k writes and 10% 8k writes, you would
589 bssplit=2k/50:4k/50,4k/90:8k/10
592 bs_unaligned If this option is given, any byte size value within bsrange
593 may be used as a block range. This typically wont work with
594 direct IO, as that normally requires sector alignment.
596 bs_is_seq_rand If this option is set, fio will use the normal read,write
597 blocksize settings as sequential,random instead. Any random
598 read or write will use the WRITE blocksize settings, and any
599 sequential read or write will use the READ blocksize setting.
601 zero_buffers If this option is given, fio will init the IO buffers to
602 all zeroes. The default is to fill them with random data.
604 refill_buffers If this option is given, fio will refill the IO buffers
605 on every submit. The default is to only fill it at init
606 time and reuse that data. Only makes sense if zero_buffers
607 isn't specified, naturally. If data verification is enabled,
608 refill_buffers is also automatically enabled.
610 scramble_buffers=bool If refill_buffers is too costly and the target is
611 using data deduplication, then setting this option will
612 slightly modify the IO buffer contents to defeat normal
613 de-dupe attempts. This is not enough to defeat more clever
614 block compression attempts, but it will stop naive dedupe of
615 blocks. Default: true.
617 buffer_compress_percentage=int If this is set, then fio will attempt to
618 provide IO buffer content (on WRITEs) that compress to
619 the specified level. Fio does this by providing a mix of
620 random data and a fixed pattern. The fixed pattern is either
621 zeroes, or the pattern specified by buffer_pattern. If the
622 pattern option is used, it might skew the compression ratio
623 slightly. Note that this is per block size unit, for file/disk
624 wide compression level that matches this setting, you'll also
625 want to set refill_buffers.
627 buffer_compress_chunk=int See buffer_compress_percentage. This
628 setting allows fio to manage how big the ranges of random
629 data and zeroed data is. Without this set, fio will
630 provide buffer_compress_percentage of blocksize random
631 data, followed by the remaining zeroed. With this set
632 to some chunk size smaller than the block size, fio can
633 alternate random and zeroed data throughout the IO
636 buffer_pattern=str If set, fio will fill the io buffers with this
637 pattern. If not set, the contents of io buffers is defined by
638 the other options related to buffer contents. The setting can
639 be any pattern of bytes, and can be prefixed with 0x for hex
640 values. It may also be a string, where the string must then
641 be wrapped with "", e.g.:
643 buffer_pattern="abcd"
647 buffer_pattern=0xdeadface
649 Also you can combine everything together in any order:
650 buffer_pattern=0xdeadface"abcd"-12
652 dedupe_percentage=int If set, fio will generate this percentage of
653 identical buffers when writing. These buffers will be
654 naturally dedupable. The contents of the buffers depend on
655 what other buffer compression settings have been set. It's
656 possible to have the individual buffers either fully
657 compressible, or not at all. This option only controls the
658 distribution of unique buffers.
660 nrfiles=int Number of files to use for this job. Defaults to 1.
662 openfiles=int Number of files to keep open at the same time. Defaults to
663 the same as nrfiles, can be set smaller to limit the number
666 file_service_type=str Defines how fio decides which file from a job to
667 service next. The following types are defined:
669 random Just choose a file at random.
671 roundrobin Round robin over open files. This
674 sequential Finish one file before moving on to
675 the next. Multiple files can still be
676 open depending on 'openfiles'.
678 zipf Use a zipfian distribution to decide what file
681 pareto Use a pareto distribution to decide what file
684 gauss Use a gaussian (normal) distribution to decide
687 For random, roundrobin, and sequential, a postfix can be
688 appended to tell fio how many I/Os to issue before switching
689 to a new file. For example, specifying
690 'file_service_type=random:8' would cause fio to issue 8 I/Os
691 before selecting a new file at random. For the non-uniform
692 distributions, a floating point postfix can be given to
693 influence how the distribution is skewed. See
694 'random_distribution' for a description of how that would work.
696 ioengine=str Defines how the job issues io to the file. The following
699 sync Basic read(2) or write(2) io. lseek(2) is
700 used to position the io location.
702 psync Basic pread(2) or pwrite(2) io.
704 vsync Basic readv(2) or writev(2) IO.
706 pvsync Basic preadv(2) or pwritev(2) IO.
708 pvsync2 Basic preadv2(2) or pwritev2(2) IO.
710 libaio Linux native asynchronous io. Note that Linux
711 may only support queued behaviour with
712 non-buffered IO (set direct=1 or buffered=0).
713 This engine defines engine specific options.
715 posixaio glibc posix asynchronous io.
717 solarisaio Solaris native asynchronous io.
719 windowsaio Windows native asynchronous io.
721 mmap File is memory mapped and data copied
722 to/from using memcpy(3).
724 splice splice(2) is used to transfer the data and
725 vmsplice(2) to transfer data from user
728 sg SCSI generic sg v3 io. May either be
729 synchronous using the SG_IO ioctl, or if
730 the target is an sg character device
731 we use read(2) and write(2) for asynchronous
734 null Doesn't transfer any data, just pretends
735 to. This is mainly used to exercise fio
736 itself and for debugging/testing purposes.
738 net Transfer over the network to given host:port.
739 Depending on the protocol used, the hostname,
740 port, listen and filename options are used to
741 specify what sort of connection to make, while
742 the protocol option determines which protocol
744 This engine defines engine specific options.
746 netsplice Like net, but uses splice/vmsplice to
747 map data and send/receive.
748 This engine defines engine specific options.
750 cpuio Doesn't transfer any data, but burns CPU
751 cycles according to the cpuload= and
752 cpucycle= options. Setting cpuload=85
753 will cause that job to do nothing but burn
754 85% of the CPU. In case of SMP machines,
755 use numjobs=<no_of_cpu> to get desired CPU
756 usage, as the cpuload only loads a single
757 CPU at the desired rate.
759 guasi The GUASI IO engine is the Generic Userspace
760 Asyncronous Syscall Interface approach
763 http://www.xmailserver.org/guasi-lib.html
765 for more info on GUASI.
767 rdma The RDMA I/O engine supports both RDMA
768 memory semantics (RDMA_WRITE/RDMA_READ) and
769 channel semantics (Send/Recv) for the
770 InfiniBand, RoCE and iWARP protocols.
772 falloc IO engine that does regular fallocate to
773 simulate data transfer as fio ioengine.
774 DDIR_READ does fallocate(,mode = keep_size,)
775 DDIR_WRITE does fallocate(,mode = 0)
776 DDIR_TRIM does fallocate(,mode = punch_hole)
778 e4defrag IO engine that does regular EXT4_IOC_MOVE_EXT
779 ioctls to simulate defragment activity in
780 request to DDIR_WRITE event
782 rbd IO engine supporting direct access to Ceph
783 Rados Block Devices (RBD) via librbd without
784 the need to use the kernel rbd driver. This
785 ioengine defines engine specific options.
787 gfapi Using Glusterfs libgfapi sync interface to
788 direct access to Glusterfs volumes without
791 gfapi_async Using Glusterfs libgfapi async interface
792 to direct access to Glusterfs volumes without
793 having to go through FUSE. This ioengine
794 defines engine specific options.
796 libhdfs Read and write through Hadoop (HDFS).
797 This engine interprets offsets a little
798 differently. In HDFS, files once created
799 cannot be modified. So random writes are not
800 possible. To imitate this, libhdfs engine
801 creates bunch of small files, and engine will
802 pick a file out of those files based on the
803 offset enerated by fio backend. Each jobs uses
804 it's own connection to HDFS.
806 mtd Read, write and erase an MTD character device
807 (e.g., /dev/mtd0). Discards are treated as
808 erases. Depending on the underlying device
809 type, the I/O may have to go in a certain
810 pattern, e.g., on NAND, writing sequentially
811 to erase blocks and discarding before
812 overwriting. The writetrim mode works well
815 pmemblk Read and write through the NVML libpmemblk
818 external Prefix to specify loading an external
819 IO engine object file. Append the engine
820 filename, eg ioengine=external:/tmp/foo.o
821 to load ioengine foo.o in /tmp.
823 iodepth=int This defines how many io units to keep in flight against
824 the file. The default is 1 for each file defined in this
825 job, can be overridden with a larger value for higher
826 concurrency. Note that increasing iodepth beyond 1 will not
827 affect synchronous ioengines (except for small degress when
828 verify_async is in use). Even async engines may impose OS
829 restrictions causing the desired depth not to be achieved.
830 This may happen on Linux when using libaio and not setting
831 direct=1, since buffered IO is not async on that OS. Keep an
832 eye on the IO depth distribution in the fio output to verify
833 that the achieved depth is as expected. Default: 1.
835 iodepth_batch_submit=int
836 iodepth_batch=int This defines how many pieces of IO to submit at once.
837 It defaults to 1 which means that we submit each IO
838 as soon as it is available, but can be raised to submit
839 bigger batches of IO at the time. If it is set to 0 the iodepth
842 iodepth_batch_complete_min=int
843 iodepth_batch_complete=int This defines how many pieces of IO to retrieve
844 at once. It defaults to 1 which means that we'll ask
845 for a minimum of 1 IO in the retrieval process from
846 the kernel. The IO retrieval will go on until we
847 hit the limit set by iodepth_low. If this variable is
848 set to 0, then fio will always check for completed
849 events before queuing more IO. This helps reduce
850 IO latency, at the cost of more retrieval system calls.
852 iodepth_batch_complete_max=int This defines maximum pieces of IO to
853 retrieve at once. This variable should be used along with
854 iodepth_batch_complete_min=int variable, specifying the range
855 of min and max amount of IO which should be retrieved. By default
856 it is equal to iodepth_batch_complete_min value.
860 iodepth_batch_complete_min=1
861 iodepth_batch_complete_max=<iodepth>
863 which means that we will retrieve at leat 1 IO and up to the
864 whole submitted queue depth. If none of IO has been completed
869 iodepth_batch_complete_min=0
870 iodepth_batch_complete_max=<iodepth>
872 which means that we can retrieve up to the whole submitted
873 queue depth, but if none of IO has been completed yet, we will
874 NOT wait and immediately exit the system call. In this example
875 we simply do polling.
877 iodepth_low=int The low water mark indicating when to start filling
878 the queue again. Defaults to the same as iodepth, meaning
879 that fio will attempt to keep the queue full at all times.
880 If iodepth is set to eg 16 and iodepth_low is set to 4, then
881 after fio has filled the queue of 16 requests, it will let
882 the depth drain down to 4 before starting to fill it again.
884 io_submit_mode=str This option controls how fio submits the IO to
885 the IO engine. The default is 'inline', which means that the
886 fio job threads submit and reap IO directly. If set to
887 'offload', the job threads will offload IO submission to a
888 dedicated pool of IO threads. This requires some coordination
889 and thus has a bit of extra overhead, especially for lower
890 queue depth IO where it can increase latencies. The benefit
891 is that fio can manage submission rates independently of
892 the device completion rates. This avoids skewed latency
893 reporting if IO gets back up on the device side (the
894 coordinated omission problem).
896 direct=bool If value is true, use non-buffered io. This is usually
897 O_DIRECT. Note that ZFS on Solaris doesn't support direct io.
898 On Windows the synchronous ioengines don't support direct io.
900 atomic=bool If value is true, attempt to use atomic direct IO. Atomic
901 writes are guaranteed to be stable once acknowledged by
902 the operating system. Only Linux supports O_ATOMIC right
905 buffered=bool If value is true, use buffered io. This is the opposite
906 of the 'direct' option. Defaults to true.
908 offset=int Start io at the given offset in the file. The data before
909 the given offset will not be touched. This effectively
910 caps the file size at real_size - offset.
912 offset_increment=int If this is provided, then the real offset becomes
913 offset + offset_increment * thread_number, where the thread
914 number is a counter that starts at 0 and is incremented for
915 each sub-job (i.e. when numjobs option is specified). This
916 option is useful if there are several jobs which are intended
917 to operate on a file in parallel disjoint segments, with
918 even spacing between the starting points.
920 number_ios=int Fio will normally perform IOs until it has exhausted the size
921 of the region set by size=, or if it exhaust the allocated
922 time (or hits an error condition). With this setting, the
923 range/size can be set independently of the number of IOs to
924 perform. When fio reaches this number, it will exit normally
925 and report status. Note that this does not extend the amount
926 of IO that will be done, it will only stop fio if this
927 condition is met before other end-of-job criteria.
929 fsync=int If writing to a file, issue a sync of the dirty data
930 for every number of blocks given. For example, if you give
931 32 as a parameter, fio will sync the file for every 32
932 writes issued. If fio is using non-buffered io, we may
933 not sync the file. The exception is the sg io engine, which
934 synchronizes the disk cache anyway.
936 fdatasync=int Like fsync= but uses fdatasync() to only sync data and not
938 In FreeBSD and Windows there is no fdatasync(), this falls back
941 sync_file_range=str:val Use sync_file_range() for every 'val' number of
942 write operations. Fio will track range of writes that
943 have happened since the last sync_file_range() call. 'str'
944 can currently be one or more of:
946 wait_before SYNC_FILE_RANGE_WAIT_BEFORE
947 write SYNC_FILE_RANGE_WRITE
948 wait_after SYNC_FILE_RANGE_WAIT_AFTER
950 So if you do sync_file_range=wait_before,write:8, fio would
951 use SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE for
952 every 8 writes. Also see the sync_file_range(2) man page.
953 This option is Linux specific.
955 overwrite=bool If true, writes to a file will always overwrite existing
956 data. If the file doesn't already exist, it will be
957 created before the write phase begins. If the file exists
958 and is large enough for the specified write phase, nothing
961 end_fsync=bool If true, fsync file contents when a write stage has completed.
963 fsync_on_close=bool If true, fio will fsync() a dirty file on close.
964 This differs from end_fsync in that it will happen on every
965 file close, not just at the end of the job.
967 rwmixread=int How large a percentage of the mix should be reads.
969 rwmixwrite=int How large a percentage of the mix should be writes. If both
970 rwmixread and rwmixwrite is given and the values do not add
971 up to 100%, the latter of the two will be used to override
972 the first. This may interfere with a given rate setting,
973 if fio is asked to limit reads or writes to a certain rate.
974 If that is the case, then the distribution may be skewed.
976 random_distribution=str:float By default, fio will use a completely uniform
977 random distribution when asked to perform random IO. Sometimes
978 it is useful to skew the distribution in specific ways,
979 ensuring that some parts of the data is more hot than others.
980 fio includes the following distribution models:
982 random Uniform random distribution
983 zipf Zipf distribution
984 pareto Pareto distribution
985 gauss Normal (guassian) distribution
986 zoned Zoned random distribution
988 When using a zipf or pareto distribution, an input value
989 is also needed to define the access pattern. For zipf, this
990 is the zipf theta. For pareto, it's the pareto power. Fio
991 includes a test program, genzipf, that can be used visualize
992 what the given input values will yield in terms of hit rates.
993 If you wanted to use zipf with a theta of 1.2, you would use
994 random_distribution=zipf:1.2 as the option. If a non-uniform
995 model is used, fio will disable use of the random map. For
996 the gauss distribution, a normal deviation is supplied as
997 a value between 0 and 100.
999 For a zoned distribution, fio supports specifying percentages
1000 of IO access that should fall within what range of the file or
1001 device. For example, given a criteria of:
1003 60% of accesses should be to the first 10%
1004 30% of accesses should be to the next 20%
1005 8% of accesses should be to to the next 30%
1006 2% of accesses should be to the next 40%
1008 we can define that through zoning of the random accesses. For
1009 the above example, the user would do:
1011 random_distribution=zoned:60/10:30/20:8/30:2/40
1013 similarly to how bssplit works for setting ranges and
1014 percentages of block sizes. Like bssplit, it's possible to
1015 specify separate zones for reads, writes, and trims. If just
1016 one set is given, it'll apply to all of them.
1018 percentage_random=int For a random workload, set how big a percentage should
1019 be random. This defaults to 100%, in which case the workload
1020 is fully random. It can be set from anywhere from 0 to 100.
1021 Setting it to 0 would make the workload fully sequential. Any
1022 setting in between will result in a random mix of sequential
1023 and random IO, at the given percentages. It is possible to
1024 set different values for reads, writes, and trim. To do so,
1025 simply use a comma separated list. See blocksize.
1027 norandommap Normally fio will cover every block of the file when doing
1028 random IO. If this option is given, fio will just get a
1029 new random offset without looking at past io history. This
1030 means that some blocks may not be read or written, and that
1031 some blocks may be read/written more than once. If this option
1032 is used with verify= and multiple blocksizes (via bsrange=),
1033 only intact blocks are verified, i.e., partially-overwritten
1036 softrandommap=bool See norandommap. If fio runs with the random block map
1037 enabled and it fails to allocate the map, if this option is
1038 set it will continue without a random block map. As coverage
1039 will not be as complete as with random maps, this option is
1040 disabled by default.
1042 random_generator=str Fio supports the following engines for generating
1043 IO offsets for random IO:
1045 tausworthe Strong 2^88 cycle random number generator
1046 lfsr Linear feedback shift register generator
1047 tausworthe64 Strong 64-bit 2^258 cycle random number
1050 Tausworthe is a strong random number generator, but it
1051 requires tracking on the side if we want to ensure that
1052 blocks are only read or written once. LFSR guarantees
1053 that we never generate the same offset twice, and it's
1054 also less computationally expensive. It's not a true
1055 random generator, however, though for IO purposes it's
1056 typically good enough. LFSR only works with single
1057 block sizes, not with workloads that use multiple block
1058 sizes. If used with such a workload, fio may read or write
1059 some blocks multiple times. The default value is tausworthe,
1060 unless the required space exceeds 2^32 blocks. If it does,
1061 then tausworthe64 is selected automatically.
1063 nice=int Run the job with the given nice value. See man nice(2).
1065 prio=int Set the io priority value of this job. Linux limits us to
1066 a positive value between 0 and 7, with 0 being the highest.
1069 prioclass=int Set the io priority class. See man ionice(1).
1071 thinktime=int Stall the job x microseconds after an io has completed before
1072 issuing the next. May be used to simulate processing being
1073 done by an application. See thinktime_blocks and
1077 Only valid if thinktime is set - pretend to spend CPU time
1078 doing something with the data received, before falling back
1079 to sleeping for the rest of the period specified by
1082 thinktime_blocks=int
1083 Only valid if thinktime is set - control how many blocks
1084 to issue, before waiting 'thinktime' usecs. If not set,
1085 defaults to 1 which will make fio wait 'thinktime' usecs
1086 after every block. This effectively makes any queue depth
1087 setting redundant, since no more than 1 IO will be queued
1088 before we have to complete it and do our thinktime. In
1089 other words, this setting effectively caps the queue depth
1090 if the latter is larger.
1092 rate=int Cap the bandwidth used by this job. The number is in bytes/sec,
1093 the normal suffix rules apply. You can use rate=500k to limit
1094 reads and writes to 500k each, or you can specify read and
1095 writes separately. Using rate=1m,500k would limit reads to
1096 1MB/sec and writes to 500KB/sec. Capping only reads or
1097 writes can be done with rate=,500k or rate=500k,. The former
1098 will only limit writes (to 500KB/sec), the latter will only
1101 rate_min=int Tell fio to do whatever it can to maintain at least this
1102 bandwidth. Failing to meet this requirement, will cause
1103 the job to exit. The same format as rate is used for
1104 read vs write separation.
1106 rate_iops=int Cap the bandwidth to this number of IOPS. Basically the same
1107 as rate, just specified independently of bandwidth. If the
1108 job is given a block size range instead of a fixed value,
1109 the smallest block size is used as the metric. The same format
1110 as rate is used for read vs write separation.
1112 rate_iops_min=int If fio doesn't meet this rate of IO, it will cause
1113 the job to exit. The same format as rate is used for read vs
1116 rate_process=str This option controls how fio manages rated IO
1117 submissions. The default is 'linear', which submits IO in a
1118 linear fashion with fixed delays between IOs that gets
1119 adjusted based on IO completion rates. If this is set to
1120 'poisson', fio will submit IO based on a more real world
1121 random request flow, known as the Poisson process
1122 (https://en.wikipedia.org/wiki/Poisson_process). The lambda
1123 will be 10^6 / IOPS for the given workload.
1125 latency_target=int If set, fio will attempt to find the max performance
1126 point that the given workload will run at while maintaining a
1127 latency below this target. The values is given in microseconds.
1128 See latency_window and latency_percentile
1130 latency_window=int Used with latency_target to specify the sample window
1131 that the job is run at varying queue depths to test the
1132 performance. The value is given in microseconds.
1134 latency_percentile=float The percentage of IOs that must fall within the
1135 criteria specified by latency_target and latency_window. If not
1136 set, this defaults to 100.0, meaning that all IOs must be equal
1137 or below to the value set by latency_target.
1139 max_latency=int If set, fio will exit the job if it exceeds this maximum
1140 latency. It will exit with an ETIME error.
1142 rate_cycle=int Average bandwidth for 'rate' and 'rate_min' over this number
1145 cpumask=int Set the CPU affinity of this job. The parameter given is a
1146 bitmask of allowed CPU's the job may run on. So if you want
1147 the allowed CPUs to be 1 and 5, you would pass the decimal
1148 value of (1 << 1 | 1 << 5), or 34. See man
1149 sched_setaffinity(2). This may not work on all supported
1150 operating systems or kernel versions. This option doesn't
1151 work well for a higher CPU count than what you can store in
1152 an integer mask, so it can only control cpus 1-32. For
1153 boxes with larger CPU counts, use cpus_allowed.
1155 cpus_allowed=str Controls the same options as cpumask, but it allows a text
1156 setting of the permitted CPUs instead. So to use CPUs 1 and
1157 5, you would specify cpus_allowed=1,5. This options also
1158 allows a range of CPUs. Say you wanted a binding to CPUs
1159 1, 5, and 8-15, you would set cpus_allowed=1,5,8-15.
1161 cpus_allowed_policy=str Set the policy of how fio distributes the CPUs
1162 specified by cpus_allowed or cpumask. Two policies are
1165 shared All jobs will share the CPU set specified.
1166 split Each job will get a unique CPU from the CPU set.
1168 'shared' is the default behaviour, if the option isn't
1169 specified. If split is specified, then fio will will assign
1170 one cpu per job. If not enough CPUs are given for the jobs
1171 listed, then fio will roundrobin the CPUs in the set.
1173 numa_cpu_nodes=str Set this job running on spcified NUMA nodes' CPUs. The
1174 arguments allow comma delimited list of cpu numbers,
1175 A-B ranges, or 'all'. Note, to enable numa options support,
1176 fio must be built on a system with libnuma-dev(el) installed.
1178 numa_mem_policy=str Set this job's memory policy and corresponding NUMA
1179 nodes. Format of the argements:
1181 `mode' is one of the following memory policy:
1182 default, prefer, bind, interleave, local
1183 For `default' and `local' memory policy, no node is
1184 needed to be specified.
1185 For `prefer', only one node is allowed.
1186 For `bind' and `interleave', it allow comma delimited
1187 list of numbers, A-B ranges, or 'all'.
1189 startdelay=time Start this job the specified number of seconds after fio
1190 has started. Only useful if the job file contains several
1191 jobs, and you want to delay starting some jobs to a certain
1194 runtime=time Tell fio to terminate processing after the specified number
1195 of seconds. It can be quite hard to determine for how long
1196 a specified job will run, so this parameter is handy to
1197 cap the total runtime to a given time.
1199 time_based If set, fio will run for the duration of the runtime
1200 specified even if the file(s) are completely read or
1201 written. It will simply loop over the same workload
1202 as many times as the runtime allows.
1204 ramp_time=time If set, fio will run the specified workload for this amount
1205 of time before logging any performance numbers. Useful for
1206 letting performance settle before logging results, thus
1207 minimizing the runtime required for stable results. Note
1208 that the ramp_time is considered lead in time for a job,
1209 thus it will increase the total runtime if a special timeout
1210 or runtime is specified.
1212 invalidate=bool Invalidate the buffer/page cache parts for this file prior
1213 to starting io. Defaults to true.
1215 sync=bool Use sync io for buffered writes. For the majority of the
1216 io engines, this means using O_SYNC.
1219 mem=str Fio can use various types of memory as the io unit buffer.
1220 The allowed values are:
1222 malloc Use memory from malloc(3) as the buffers.
1224 shm Use shared memory as the buffers. Allocated
1227 shmhuge Same as shm, but use huge pages as backing.
1229 mmap Use mmap to allocate buffers. May either be
1230 anonymous memory, or can be file backed if
1231 a filename is given after the option. The
1232 format is mem=mmap:/path/to/file.
1234 mmaphuge Use a memory mapped huge file as the buffer
1235 backing. Append filename after mmaphuge, ala
1236 mem=mmaphuge:/hugetlbfs/file
1238 mmapshared Same as mmap, but use a MMAP_SHARED
1241 The area allocated is a function of the maximum allowed
1242 bs size for the job, multiplied by the io depth given. Note
1243 that for shmhuge and mmaphuge to work, the system must have
1244 free huge pages allocated. This can normally be checked
1245 and set by reading/writing /proc/sys/vm/nr_hugepages on a
1246 Linux system. Fio assumes a huge page is 4MB in size. So
1247 to calculate the number of huge pages you need for a given
1248 job file, add up the io depth of all jobs (normally one unless
1249 iodepth= is used) and multiply by the maximum bs set. Then
1250 divide that number by the huge page size. You can see the
1251 size of the huge pages in /proc/meminfo. If no huge pages
1252 are allocated by having a non-zero number in nr_hugepages,
1253 using mmaphuge or shmhuge will fail. Also see hugepage-size.
1255 mmaphuge also needs to have hugetlbfs mounted and the file
1256 location should point there. So if it's mounted in /huge,
1257 you would use mem=mmaphuge:/huge/somefile.
1259 iomem_align=int This indiciates the memory alignment of the IO memory buffers.
1260 Note that the given alignment is applied to the first IO unit
1261 buffer, if using iodepth the alignment of the following buffers
1262 are given by the bs used. In other words, if using a bs that is
1263 a multiple of the page sized in the system, all buffers will
1264 be aligned to this value. If using a bs that is not page
1265 aligned, the alignment of subsequent IO memory buffers is the
1266 sum of the iomem_align and bs used.
1269 Defines the size of a huge page. Must at least be equal
1270 to the system setting, see /proc/meminfo. Defaults to 4MB.
1271 Should probably always be a multiple of megabytes, so using
1272 hugepage-size=Xm is the preferred way to set this to avoid
1273 setting a non-pow-2 bad value.
1275 exitall When one job finishes, terminate the rest. The default is
1276 to wait for each job to finish, sometimes that is not the
1279 exitall_on_error When one job finishes in error, terminate the rest. The
1280 default is to wait for each job to finish.
1282 bwavgtime=int Average the calculated bandwidth over the given time. Value
1283 is specified in milliseconds. If the job also does bandwidth
1284 logging through 'write_bw_log', then the minimum of this option
1285 and 'log_avg_msec' will be used. Default: 500ms.
1287 iopsavgtime=int Average the calculated IOPS over the given time. Value
1288 is specified in milliseconds. If the job also does IOPS logging
1289 through 'write_iops_log', then the minimum of this option and
1290 'log_avg_msec' will be used. Default: 500ms.
1292 create_serialize=bool If true, serialize the file creating for the jobs.
1293 This may be handy to avoid interleaving of data
1294 files, which may greatly depend on the filesystem
1295 used and even the number of processors in the system.
1297 create_fsync=bool fsync the data file after creation. This is the
1300 create_on_open=bool Don't pre-setup the files for IO, just create open()
1301 when it's time to do IO to that file.
1303 create_only=bool If true, fio will only run the setup phase of the job.
1304 If files need to be laid out or updated on disk, only
1305 that will be done. The actual job contents are not
1308 allow_file_create=bool If true, fio is permitted to create files as part
1309 of its workload. This is the default behavior. If this
1310 option is false, then fio will error out if the files it
1311 needs to use don't already exist. Default: true.
1313 allow_mounted_write=bool If this isn't set, fio will abort jobs that
1314 are destructive (eg that write) to what appears to be a
1315 mounted device or partition. This should help catch creating
1316 inadvertently destructive tests, not realizing that the test
1317 will destroy data on the mounted file system. Default: false.
1319 pre_read=bool If this is given, files will be pre-read into memory before
1320 starting the given IO operation. This will also clear
1321 the 'invalidate' flag, since it is pointless to pre-read
1322 and then drop the cache. This will only work for IO engines
1323 that are seekable, since they allow you to read the same data
1324 multiple times. Thus it will not work on eg network or splice
1327 unlink=bool Unlink the job files when done. Not the default, as repeated
1328 runs of that job would then waste time recreating the file
1329 set again and again.
1331 loops=int Run the specified number of iterations of this job. Used
1332 to repeat the same workload a given number of times. Defaults
1335 verify_only Do not perform specified workload---only verify data still
1336 matches previous invocation of this workload. This option
1337 allows one to check data multiple times at a later date
1338 without overwriting it. This option makes sense only for
1339 workloads that write data, and does not support workloads
1340 with the time_based option set.
1342 do_verify=bool Run the verify phase after a write phase. Only makes sense if
1343 verify is set. Defaults to 1.
1345 verify=str If writing to a file, fio can verify the file contents
1346 after each iteration of the job. Each verification method also implies
1347 verification of special header, which is written to the beginning of
1348 each block. This header also includes meta information, like offset
1349 of the block, block number, timestamp when block was written, etc.
1350 verify=str can be combined with verify_pattern=str option.
1351 The allowed values are:
1353 md5 Use an md5 sum of the data area and store
1354 it in the header of each block.
1356 crc64 Use an experimental crc64 sum of the data
1357 area and store it in the header of each
1360 crc32c Use a crc32c sum of the data area and store
1361 it in the header of each block.
1363 crc32c-intel Use hardware assisted crc32c calcuation
1364 provided on SSE4.2 enabled processors. Falls
1365 back to regular software crc32c, if not
1366 supported by the system.
1368 crc32 Use a crc32 sum of the data area and store
1369 it in the header of each block.
1371 crc16 Use a crc16 sum of the data area and store
1372 it in the header of each block.
1374 crc7 Use a crc7 sum of the data area and store
1375 it in the header of each block.
1377 xxhash Use xxhash as the checksum function. Generally
1378 the fastest software checksum that fio
1381 sha512 Use sha512 as the checksum function.
1383 sha256 Use sha256 as the checksum function.
1385 sha1 Use optimized sha1 as the checksum function.
1387 meta This option is deprecated, since now meta information is
1388 included in generic verification header and meta verification
1389 happens by default. For detailed information see the description
1390 of the verify=str setting. This option is kept because of
1391 compatibility's sake with old configurations. Do not use it.
1393 pattern Verify a strict pattern. Normally fio includes
1394 a header with some basic information and
1395 checksumming, but if this option is set, only
1396 the specific pattern set with 'verify_pattern'
1399 null Only pretend to verify. Useful for testing
1400 internals with ioengine=null, not for much
1403 This option can be used for repeated burn-in tests of a
1404 system to make sure that the written data is also
1405 correctly read back. If the data direction given is
1406 a read or random read, fio will assume that it should
1407 verify a previously written file. If the data direction
1408 includes any form of write, the verify will be of the
1411 verifysort=bool If set, fio will sort written verify blocks when it deems
1412 it faster to read them back in a sorted manner. This is
1413 often the case when overwriting an existing file, since
1414 the blocks are already laid out in the file system. You
1415 can ignore this option unless doing huge amounts of really
1416 fast IO where the red-black tree sorting CPU time becomes
1419 verify_offset=int Swap the verification header with data somewhere else
1420 in the block before writing. Its swapped back before
1423 verify_interval=int Write the verification header at a finer granularity
1424 than the blocksize. It will be written for chunks the
1425 size of header_interval. blocksize should divide this
1428 verify_pattern=str If set, fio will fill the io buffers with this
1429 pattern. Fio defaults to filling with totally random
1430 bytes, but sometimes it's interesting to fill with a known
1431 pattern for io verification purposes. Depending on the
1432 width of the pattern, fio will fill 1/2/3/4 bytes of the
1433 buffer at the time(it can be either a decimal or a hex number).
1434 The verify_pattern if larger than a 32-bit quantity has to
1435 be a hex number that starts with either "0x" or "0X". Use
1436 with verify=str. Also, verify_pattern supports %o format,
1437 which means that for each block offset will be written and
1438 then verifyied back, e.g.:
1442 Or use combination of everything:
1443 verify_pattern=0xff%o"abcd"-12
1445 verify_fatal=bool Normally fio will keep checking the entire contents
1446 before quitting on a block verification failure. If this
1447 option is set, fio will exit the job on the first observed
1450 verify_dump=bool If set, dump the contents of both the original data
1451 block and the data block we read off disk to files. This
1452 allows later analysis to inspect just what kind of data
1453 corruption occurred. Off by default.
1455 verify_async=int Fio will normally verify IO inline from the submitting
1456 thread. This option takes an integer describing how many
1457 async offload threads to create for IO verification instead,
1458 causing fio to offload the duty of verifying IO contents
1459 to one or more separate threads. If using this offload
1460 option, even sync IO engines can benefit from using an
1461 iodepth setting higher than 1, as it allows them to have
1462 IO in flight while verifies are running.
1464 verify_async_cpus=str Tell fio to set the given CPU affinity on the
1465 async IO verification threads. See cpus_allowed for the
1468 verify_backlog=int Fio will normally verify the written contents of a
1469 job that utilizes verify once that job has completed. In
1470 other words, everything is written then everything is read
1471 back and verified. You may want to verify continually
1472 instead for a variety of reasons. Fio stores the meta data
1473 associated with an IO block in memory, so for large
1474 verify workloads, quite a bit of memory would be used up
1475 holding this meta data. If this option is enabled, fio
1476 will write only N blocks before verifying these blocks.
1478 verify_backlog_batch=int Control how many blocks fio will verify
1479 if verify_backlog is set. If not set, will default to
1480 the value of verify_backlog (meaning the entire queue
1481 is read back and verified). If verify_backlog_batch is
1482 less than verify_backlog then not all blocks will be verified,
1483 if verify_backlog_batch is larger than verify_backlog, some
1484 blocks will be verified more than once.
1486 verify_state_save=bool When a job exits during the write phase of a verify
1487 workload, save its current state. This allows fio to replay
1488 up until that point, if the verify state is loaded for the
1489 verify read phase. The format of the filename is, roughly,
1490 <type>-<jobname>-<jobindex>-verify.state. <type> is "local"
1491 for a local run, "sock" for a client/server socket connection,
1492 and "ip" (192.168.0.1, for instance) for a networked
1493 client/server connection.
1495 verify_state_load=bool If a verify termination trigger was used, fio stores
1496 the current write state of each thread. This can be used at
1497 verification time so that fio knows how far it should verify.
1498 Without this information, fio will run a full verification
1499 pass, according to the settings in the job file used.
1502 wait_for_previous Wait for preceding jobs in the job file to exit, before
1503 starting this one. Can be used to insert serialization
1504 points in the job file. A stone wall also implies starting
1505 a new reporting group.
1507 new_group Start a new reporting group. See: group_reporting.
1509 numjobs=int Create the specified number of clones of this job. May be
1510 used to setup a larger number of threads/processes doing
1511 the same thing. Each thread is reported separately; to see
1512 statistics for all clones as a whole, use group_reporting in
1513 conjunction with new_group.
1515 group_reporting It may sometimes be interesting to display statistics for
1516 groups of jobs as a whole instead of for each individual job.
1517 This is especially true if 'numjobs' is used; looking at
1518 individual thread/process output quickly becomes unwieldy.
1519 To see the final report per-group instead of per-job, use
1520 'group_reporting'. Jobs in a file will be part of the same
1521 reporting group, unless if separated by a stonewall, or by
1524 thread fio defaults to forking jobs, however if this option is
1525 given, fio will use pthread_create(3) to create threads
1528 zonesize=int Divide a file into zones of the specified size. See zoneskip.
1530 zoneskip=int Skip the specified number of bytes when zonesize data has
1531 been read. The two zone options can be used to only do
1532 io on zones of a file.
1534 write_iolog=str Write the issued io patterns to the specified file. See
1535 read_iolog. Specify a separate file for each job, otherwise
1536 the iologs will be interspersed and the file may be corrupt.
1538 read_iolog=str Open an iolog with the specified file name and replay the
1539 io patterns it contains. This can be used to store a
1540 workload and replay it sometime later. The iolog given
1541 may also be a blktrace binary file, which allows fio
1542 to replay a workload captured by blktrace. See blktrace
1543 for how to capture such logging data. For blktrace replay,
1544 the file needs to be turned into a blkparse binary data
1545 file first (blkparse <device> -o /dev/null -d file_for_fio.bin).
1547 replay_no_stall=int When replaying I/O with read_iolog the default behavior
1548 is to attempt to respect the time stamps within the log and
1549 replay them with the appropriate delay between IOPS. By
1550 setting this variable fio will not respect the timestamps and
1551 attempt to replay them as fast as possible while still
1552 respecting ordering. The result is the same I/O pattern to a
1553 given device, but different timings.
1555 replay_redirect=str While replaying I/O patterns using read_iolog the
1556 default behavior is to replay the IOPS onto the major/minor
1557 device that each IOP was recorded from. This is sometimes
1558 undesirable because on a different machine those major/minor
1559 numbers can map to a different device. Changing hardware on
1560 the same system can also result in a different major/minor
1561 mapping. Replay_redirect causes all IOPS to be replayed onto
1562 the single specified device regardless of the device it was
1563 recorded from. i.e. replay_redirect=/dev/sdc would cause all
1564 IO in the blktrace to be replayed onto /dev/sdc. This means
1565 multiple devices will be replayed onto a single, if the trace
1566 contains multiple devices. If you want multiple devices to be
1567 replayed concurrently to multiple redirected devices you must
1568 blkparse your trace into separate traces and replay them with
1569 independent fio invocations. Unfortuantely this also breaks
1570 the strict time ordering between multiple device accesses.
1572 replay_align=int Force alignment of IO offsets and lengths in a trace
1573 to this power of 2 value.
1575 replay_scale=int Scale sector offsets down by this factor when
1578 per_job_logs=bool If set, this generates bw/clat/iops log with per
1579 file private filenames. If not set, jobs with identical names
1580 will share the log filename. Default: true.
1582 write_bw_log=str If given, write a bandwidth log of the jobs in this job
1583 file. Can be used to store data of the bandwidth of the
1584 jobs in their lifetime. The included fio_generate_plots
1585 script uses gnuplot to turn these text files into nice
1586 graphs. See write_lat_log for behaviour of given
1587 filename. For this option, the suffix is _bw.x.log, where
1588 x is the index of the job (1..N, where N is the number of
1589 jobs). If 'per_job_logs' is false, then the filename will not
1590 include the job index. See 'Log File Formats'.
1592 write_lat_log=str Same as write_bw_log, except that this option stores io
1593 submission, completion, and total latencies instead. If no
1594 filename is given with this option, the default filename of
1595 "jobname_type.log" is used. Even if the filename is given,
1596 fio will still append the type of log. So if one specifies
1600 The actual log names will be foo_slat.x.log, foo_clat.x.log,
1601 and foo_lat.x.log, where x is the index of the job (1..N,
1602 where N is the number of jobs). This helps fio_generate_plot
1603 fine the logs automatically. If 'per_job_logs' is false, then
1604 the filename will not include the job index. See 'Log File
1607 write_iops_log=str Same as write_bw_log, but writes IOPS. If no filename is
1608 given with this option, the default filename of
1609 "jobname_type.x.log" is used,where x is the index of the job
1610 (1..N, where N is the number of jobs). Even if the filename
1611 is given, fio will still append the type of log. If
1612 'per_job_logs' is false, then the filename will not include
1613 the job index. See 'Log File Formats'.
1615 log_avg_msec=int By default, fio will log an entry in the iops, latency,
1616 or bw log for every IO that completes. When writing to the
1617 disk log, that can quickly grow to a very large size. Setting
1618 this option makes fio average the each log entry over the
1619 specified period of time, reducing the resolution of the log.
1620 See log_max_value as well. Defaults to 0, logging all entries.
1622 log_max_value=bool If log_avg_msec is set, fio logs the average over that
1623 window. If you instead want to log the maximum value, set this
1624 option to 1. Defaults to 0, meaning that averaged values are
1627 log_offset=int If this is set, the iolog options will include the byte
1628 offset for the IO entry as well as the other data values.
1630 log_compression=int If this is set, fio will compress the IO logs as
1631 it goes, to keep the memory footprint lower. When a log
1632 reaches the specified size, that chunk is removed and
1633 compressed in the background. Given that IO logs are
1634 fairly highly compressible, this yields a nice memory
1635 savings for longer runs. The downside is that the
1636 compression will consume some background CPU cycles, so
1637 it may impact the run. This, however, is also true if
1638 the logging ends up consuming most of the system memory.
1639 So pick your poison. The IO logs are saved normally at the
1640 end of a run, by decompressing the chunks and storing them
1641 in the specified log file. This feature depends on the
1642 availability of zlib.
1644 log_compression_cpus=str Define the set of CPUs that are allowed to
1645 handle online log compression for the IO jobs. This can
1646 provide better isolation between performance sensitive jobs,
1647 and background compression work.
1649 log_store_compressed=bool If set, fio will store the log files in a
1650 compressed format. They can be decompressed with fio, using
1651 the --inflate-log command line parameter. The files will be
1652 stored with a .fz suffix.
1654 block_error_percentiles=bool If set, record errors in trim block-sized
1655 units from writes and trims and output a histogram of
1656 how many trims it took to get to errors, and what kind
1657 of error was encountered.
1659 lockmem=int Pin down the specified amount of memory with mlock(2). Can
1660 potentially be used instead of removing memory or booting
1661 with less memory to simulate a smaller amount of memory.
1662 The amount specified is per worker.
1664 exec_prerun=str Before running this job, issue the command specified
1665 through system(3). Output is redirected in a file called
1668 exec_postrun=str After the job completes, issue the command specified
1669 though system(3). Output is redirected in a file called
1670 jobname.postrun.txt.
1672 ioscheduler=str Attempt to switch the device hosting the file to the specified
1673 io scheduler before running.
1675 disk_util=bool Generate disk utilization statistics, if the platform
1676 supports it. Defaults to on.
1678 disable_lat=bool Disable measurements of total latency numbers. Useful
1679 only for cutting back the number of calls to gettimeofday,
1680 as that does impact performance at really high IOPS rates.
1681 Note that to really get rid of a large amount of these
1682 calls, this option must be used with disable_slat and
1685 disable_clat=bool Disable measurements of completion latency numbers. See
1688 disable_slat=bool Disable measurements of submission latency numbers. See
1691 disable_bw=bool Disable measurements of throughput/bandwidth numbers. See
1694 clat_percentiles=bool Enable the reporting of percentiles of
1695 completion latencies.
1697 percentile_list=float_list Overwrite the default list of percentiles
1698 for completion latencies and the block error histogram.
1699 Each number is a floating number in the range (0,100],
1700 and the maximum length of the list is 20. Use ':'
1701 to separate the numbers, and list the numbers in ascending
1702 order. For example, --percentile_list=99.5:99.9 will cause
1703 fio to report the values of completion latency below which
1704 99.5% and 99.9% of the observed latencies fell, respectively.
1706 clocksource=str Use the given clocksource as the base of timing. The
1707 supported options are:
1709 gettimeofday gettimeofday(2)
1711 clock_gettime clock_gettime(2)
1713 cpu Internal CPU clock source
1715 cpu is the preferred clocksource if it is reliable, as it
1716 is very fast (and fio is heavy on time calls). Fio will
1717 automatically use this clocksource if it's supported and
1718 considered reliable on the system it is running on, unless
1719 another clocksource is specifically set. For x86/x86-64 CPUs,
1720 this means supporting TSC Invariant.
1722 gtod_reduce=bool Enable all of the gettimeofday() reducing options
1723 (disable_clat, disable_slat, disable_bw) plus reduce
1724 precision of the timeout somewhat to really shrink
1725 the gettimeofday() call count. With this option enabled,
1726 we only do about 0.4% of the gtod() calls we would have
1727 done if all time keeping was enabled.
1729 gtod_cpu=int Sometimes it's cheaper to dedicate a single thread of
1730 execution to just getting the current time. Fio (and
1731 databases, for instance) are very intensive on gettimeofday()
1732 calls. With this option, you can set one CPU aside for
1733 doing nothing but logging current time to a shared memory
1734 location. Then the other threads/processes that run IO
1735 workloads need only copy that segment, instead of entering
1736 the kernel with a gettimeofday() call. The CPU set aside
1737 for doing these time calls will be excluded from other
1738 uses. Fio will manually clear it from the CPU mask of other
1741 continue_on_error=str Normally fio will exit the job on the first observed
1742 failure. If this option is set, fio will continue the job when
1743 there is a 'non-fatal error' (EIO or EILSEQ) until the runtime
1744 is exceeded or the I/O size specified is completed. If this
1745 option is used, there are two more stats that are appended,
1746 the total error count and the first error. The error field
1747 given in the stats is the first error that was hit during the
1750 The allowed values are:
1752 none Exit on any IO or verify errors.
1754 read Continue on read errors, exit on all others.
1756 write Continue on write errors, exit on all others.
1758 io Continue on any IO error, exit on all others.
1760 verify Continue on verify errors, exit on all others.
1762 all Continue on all errors.
1764 0 Backward-compatible alias for 'none'.
1766 1 Backward-compatible alias for 'all'.
1768 ignore_error=str Sometimes you want to ignore some errors during test
1769 in that case you can specify error list for each error type.
1770 ignore_error=READ_ERR_LIST,WRITE_ERR_LIST,VERIFY_ERR_LIST
1771 errors for given error type is separated with ':'. Error
1772 may be symbol ('ENOSPC', 'ENOMEM') or integer.
1774 ignore_error=EAGAIN,ENOSPC:122
1775 This option will ignore EAGAIN from READ, and ENOSPC and
1776 122(EDQUOT) from WRITE.
1778 error_dump=bool If set dump every error even if it is non fatal, true
1779 by default. If disabled only fatal error will be dumped
1781 cgroup=str Add job to this control group. If it doesn't exist, it will
1782 be created. The system must have a mounted cgroup blkio
1783 mount point for this to work. If your system doesn't have it
1784 mounted, you can do so with:
1786 # mount -t cgroup -o blkio none /cgroup
1788 cgroup_weight=int Set the weight of the cgroup to this value. See
1789 the documentation that comes with the kernel, allowed values
1790 are in the range of 100..1000.
1792 cgroup_nodelete=bool Normally fio will delete the cgroups it has created after
1793 the job completion. To override this behavior and to leave
1794 cgroups around after the job completion, set cgroup_nodelete=1.
1795 This can be useful if one wants to inspect various cgroup
1796 files after job completion. Default: false
1798 uid=int Instead of running as the invoking user, set the user ID to
1799 this value before the thread/process does any work.
1801 gid=int Set group ID, see uid.
1803 flow_id=int The ID of the flow. If not specified, it defaults to being a
1804 global flow. See flow.
1806 flow=int Weight in token-based flow control. If this value is used, then
1807 there is a 'flow counter' which is used to regulate the
1808 proportion of activity between two or more jobs. fio attempts
1809 to keep this flow counter near zero. The 'flow' parameter
1810 stands for how much should be added or subtracted to the flow
1811 counter on each iteration of the main I/O loop. That is, if
1812 one job has flow=8 and another job has flow=-1, then there
1813 will be a roughly 1:8 ratio in how much one runs vs the other.
1815 flow_watermark=int The maximum value that the absolute value of the flow
1816 counter is allowed to reach before the job must wait for a
1817 lower value of the counter.
1819 flow_sleep=int The period of time, in microseconds, to wait after the flow
1820 watermark has been exceeded before retrying operations
1822 In addition, there are some parameters which are only valid when a specific
1823 ioengine is in use. These are used identically to normal parameters, with the
1824 caveat that when used on the command line, they must come after the ioengine
1825 that defines them is selected.
1827 [libaio] userspace_reap Normally, with the libaio engine in use, fio will use
1828 the io_getevents system call to reap newly returned events.
1829 With this flag turned on, the AIO ring will be read directly
1830 from user-space to reap events. The reaping mode is only
1831 enabled when polling for a minimum of 0 events (eg when
1832 iodepth_batch_complete=0).
1834 [psyncv2] hipri Set RWF_HIPRI on IO, indicating to the kernel that
1835 it's of higher priority than normal.
1837 [cpu] cpuload=int Attempt to use the specified percentage of CPU cycles.
1839 [cpu] cpuchunks=int Split the load into cycles of the given time. In
1842 [cpu] exit_on_io_done=bool Detect when IO threads are done, then exit.
1844 [netsplice] hostname=str
1845 [net] hostname=str The host name or IP address to use for TCP or UDP based IO.
1846 If the job is a TCP listener or UDP reader, the hostname is not
1847 used and must be omitted unless it is a valid UDP multicast
1849 [libhdfs] namenode=str The host name or IP address of a HDFS cluster namenode to contact.
1851 [netsplice] port=int
1852 [net] port=int The TCP or UDP port to bind to or connect to. If this is used
1853 with numjobs to spawn multiple instances of the same job type, then this will
1854 be the starting port number since fio will use a range of ports.
1855 [libhdfs] port=int the listening port of the HFDS cluster namenode.
1857 [netsplice] interface=str
1858 [net] interface=str The IP address of the network interface used to send or
1859 receive UDP multicast
1862 [net] ttl=int Time-to-live value for outgoing UDP multicast packets.
1865 [netsplice] nodelay=bool
1866 [net] nodelay=bool Set TCP_NODELAY on TCP connections.
1868 [netsplice] protocol=str
1869 [netsplice] proto=str
1871 [net] proto=str The network protocol to use. Accepted values are:
1873 tcp Transmission control protocol
1874 tcpv6 Transmission control protocol V6
1875 udp User datagram protocol
1876 udpv6 User datagram protocol V6
1877 unix UNIX domain socket
1879 When the protocol is TCP or UDP, the port must also be given,
1880 as well as the hostname if the job is a TCP listener or UDP
1881 reader. For unix sockets, the normal filename option should be
1882 used and the port is invalid.
1884 [net] listen For TCP network connections, tell fio to listen for incoming
1885 connections rather than initiating an outgoing connection. The
1886 hostname must be omitted if this option is used.
1888 [net] pingpong Normaly a network writer will just continue writing data, and
1889 a network reader will just consume packages. If pingpong=1
1890 is set, a writer will send its normal payload to the reader,
1891 then wait for the reader to send the same payload back. This
1892 allows fio to measure network latencies. The submission
1893 and completion latencies then measure local time spent
1894 sending or receiving, and the completion latency measures
1895 how long it took for the other end to receive and send back.
1896 For UDP multicast traffic pingpong=1 should only be set for a
1897 single reader when multiple readers are listening to the same
1900 [net] window_size Set the desired socket buffer size for the connection.
1902 [net] mss Set the TCP maximum segment size (TCP_MAXSEG).
1904 [e4defrag] donorname=str
1905 File will be used as a block donor(swap extents between files)
1906 [e4defrag] inplace=int
1907 Configure donor file blocks allocation strategy
1908 0(default): Preallocate donor's file on init
1909 1 : allocate space immidietly inside defragment event,
1910 and free right after event
1912 [rbd] clustername=str Specifies the name of the Ceph cluster.
1913 [rbd] rbdname=str Specifies the name of the RBD.
1914 [rbd] pool=str Specifies the naem of the Ceph pool containing RBD.
1915 [rbd] clientname=str Specifies the username (without the 'client.' prefix)
1916 used to access the Ceph cluster. If the clustername is
1917 specified, the clientmae shall be the full type.id
1918 string. If no type. prefix is given, fio will add
1919 'client.' by default.
1921 [mtd] skip_bad=bool Skip operations against known bad blocks.
1923 [libhdfs] hdfsdirectory libhdfs will create chunk in this HDFS directory
1924 [libhdfs] chunck_size the size of the chunck to use for each file.
1927 6.0 Interpreting the output
1928 ---------------------------
1930 fio spits out a lot of output. While running, fio will display the
1931 status of the jobs created. An example of that would be:
1933 Threads: 1: [_r] [24.8% done] [ 13509/ 8334 kb/s] [eta 00h:01m:31s]
1935 The characters inside the square brackets denote the current status of
1936 each thread. The possible values (in typical life cycle order) are:
1940 P Thread setup, but not started.
1942 I Thread initialized, waiting or generating necessary data.
1943 p Thread running pre-reading file(s).
1944 R Running, doing sequential reads.
1945 r Running, doing random reads.
1946 W Running, doing sequential writes.
1947 w Running, doing random writes.
1948 M Running, doing mixed sequential reads/writes.
1949 m Running, doing mixed random reads/writes.
1950 F Running, currently waiting for fsync()
1951 f Running, finishing up (writing IO logs, etc)
1952 V Running, doing verification of written data.
1953 E Thread exited, not reaped by main thread yet.
1955 X Thread reaped, exited with an error.
1956 K Thread reaped, exited due to signal.
1958 Fio will condense the thread string as not to take up more space on the
1959 command line as is needed. For instance, if you have 10 readers and 10
1960 writers running, the output would look like this:
1962 Jobs: 20 (f=20): [R(10),W(10)] [4.0% done] [2103MB/0KB/0KB /s] [538K/0/0 iops] [eta 57m:36s]
1964 Fio will still maintain the ordering, though. So the above means that jobs
1965 1..10 are readers, and 11..20 are writers.
1967 The other values are fairly self explanatory - number of threads
1968 currently running and doing io, rate of io since last check (read speed
1969 listed first, then write speed), and the estimated completion percentage
1970 and time for the running group. It's impossible to estimate runtime of
1971 the following groups (if any). Note that the string is displayed in order,
1972 so it's possible to tell which of the jobs are currently doing what. The
1973 first character is the first job defined in the job file, and so forth.
1975 When fio is done (or interrupted by ctrl-c), it will show the data for
1976 each thread, group of threads, and disks in that order. For each data
1977 direction, the output looks like:
1979 Client1 (g=0): err= 0:
1980 write: io= 32MB, bw= 666KB/s, iops=89 , runt= 50320msec
1981 slat (msec): min= 0, max= 136, avg= 0.03, stdev= 1.92
1982 clat (msec): min= 0, max= 631, avg=48.50, stdev=86.82
1983 bw (KB/s) : min= 0, max= 1196, per=51.00%, avg=664.02, stdev=681.68
1984 cpu : usr=1.49%, sys=0.25%, ctx=7969, majf=0, minf=17
1985 IO depths : 1=0.1%, 2=0.3%, 4=0.5%, 8=99.0%, 16=0.0%, 32=0.0%, >32=0.0%
1986 submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
1987 complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
1988 issued r/w: total=0/32768, short=0/0
1989 lat (msec): 2=1.6%, 4=0.0%, 10=3.2%, 20=12.8%, 50=38.4%, 100=24.8%,
1990 lat (msec): 250=15.2%, 500=0.0%, 750=0.0%, 1000=0.0%, >=2048=0.0%
1992 The client number is printed, along with the group id and error of that
1993 thread. Below is the io statistics, here for writes. In the order listed,
1996 io= Number of megabytes io performed
1997 bw= Average bandwidth rate
1998 iops= Average IOs performed per second
1999 runt= The runtime of that thread
2000 slat= Submission latency (avg being the average, stdev being the
2001 standard deviation). This is the time it took to submit
2002 the io. For sync io, the slat is really the completion
2003 latency, since queue/complete is one operation there. This
2004 value can be in milliseconds or microseconds, fio will choose
2005 the most appropriate base and print that. In the example
2006 above, milliseconds is the best scale. Note: in --minimal mode
2007 latencies are always expressed in microseconds.
2008 clat= Completion latency. Same names as slat, this denotes the
2009 time from submission to completion of the io pieces. For
2010 sync io, clat will usually be equal (or very close) to 0,
2011 as the time from submit to complete is basically just
2012 CPU time (io has already been done, see slat explanation).
2013 bw= Bandwidth. Same names as the xlat stats, but also includes
2014 an approximate percentage of total aggregate bandwidth
2015 this thread received in this group. This last value is
2016 only really useful if the threads in this group are on the
2017 same disk, since they are then competing for disk access.
2018 cpu= CPU usage. User and system time, along with the number
2019 of context switches this thread went through, usage of
2020 system and user time, and finally the number of major
2021 and minor page faults. The CPU utilization numbers are
2022 averages for the jobs in that reporting group, while the
2023 context and fault counters are summed.
2024 IO depths= The distribution of io depths over the job life time. The
2025 numbers are divided into powers of 2, so for example the
2026 16= entries includes depths up to that value but higher
2027 than the previous entry. In other words, it covers the
2028 range from 16 to 31.
2029 IO submit= How many pieces of IO were submitting in a single submit
2030 call. Each entry denotes that amount and below, until
2031 the previous entry - eg, 8=100% mean that we submitted
2032 anywhere in between 5-8 ios per submit call.
2033 IO complete= Like the above submit number, but for completions instead.
2034 IO issued= The number of read/write requests issued, and how many
2036 IO latencies= The distribution of IO completion latencies. This is the
2037 time from when IO leaves fio and when it gets completed.
2038 The numbers follow the same pattern as the IO depths,
2039 meaning that 2=1.6% means that 1.6% of the IO completed
2040 within 2 msecs, 20=12.8% means that 12.8% of the IO
2041 took more than 10 msecs, but less than (or equal to) 20 msecs.
2043 After each client has been listed, the group statistics are printed. They
2044 will look like this:
2046 Run status group 0 (all jobs):
2047 READ: io=64MB, aggrb=22178, minb=11355, maxb=11814, mint=2840msec, maxt=2955msec
2048 WRITE: io=64MB, aggrb=1302, minb=666, maxb=669, mint=50093msec, maxt=50320msec
2050 For each data direction, it prints:
2052 io= Number of megabytes io performed.
2053 aggrb= Aggregate bandwidth of threads in this group.
2054 minb= The minimum average bandwidth a thread saw.
2055 maxb= The maximum average bandwidth a thread saw.
2056 mint= The smallest runtime of the threads in that group.
2057 maxt= The longest runtime of the threads in that group.
2059 And finally, the disk statistics are printed. They will look like this:
2061 Disk stats (read/write):
2062 sda: ios=16398/16511, merge=30/162, ticks=6853/819634, in_queue=826487, util=100.00%
2064 Each value is printed for both reads and writes, with reads first. The
2067 ios= Number of ios performed by all groups.
2068 merge= Number of merges io the io scheduler.
2069 ticks= Number of ticks we kept the disk busy.
2070 io_queue= Total time spent in the disk queue.
2071 util= The disk utilization. A value of 100% means we kept the disk
2072 busy constantly, 50% would be a disk idling half of the time.
2074 It is also possible to get fio to dump the current output while it is
2075 running, without terminating the job. To do that, send fio the USR1 signal.
2076 You can also get regularly timed dumps by using the --status-interval
2077 parameter, or by creating a file in /tmp named fio-dump-status. If fio
2078 sees this file, it will unlink it and dump the current output status.
2084 For scripted usage where you typically want to generate tables or graphs
2085 of the results, fio can output the results in a semicolon separated format.
2086 The format is one long line of values, such as:
2088 2;card0;0;0;7139336;121836;60004;1;10109;27.932460;116.933948;220;126861;3495.446807;1085.368601;226;126864;3523.635629;1089.012448;24063;99944;50.275485%;59818.274627;5540.657370;7155060;122104;60004;1;8338;29.086342;117.839068;388;128077;5032.488518;1234.785715;391;128085;5061.839412;1236.909129;23436;100928;50.287926%;59964.832030;5644.844189;14.595833%;19.394167%;123706;0;7313;0.1%;0.1%;0.1%;0.1%;0.1%;0.1%;100.0%;0.00%;0.00%;0.00%;0.00%;0.00%;0.00%;0.01%;0.02%;0.05%;0.16%;6.04%;40.40%;52.68%;0.64%;0.01%;0.00%;0.01%;0.00%;0.00%;0.00%;0.00%;0.00%
2089 A description of this job goes here.
2091 The job description (if provided) follows on a second line.
2093 To enable terse output, use the --minimal command line option. The first
2094 value is the version of the terse output format. If the output has to
2095 be changed for some reason, this number will be incremented by 1 to
2096 signify that change.
2098 Split up, the format is as follows:
2100 terse version, fio version, jobname, groupid, error
2102 Total IO (KB), bandwidth (KB/sec), IOPS, runtime (msec)
2103 Submission latency: min, max, mean, stdev (usec)
2104 Completion latency: min, max, mean, stdev (usec)
2105 Completion latency percentiles: 20 fields (see below)
2106 Total latency: min, max, mean, stdev (usec)
2107 Bw (KB/s): min, max, aggregate percentage of total, mean, stdev
2109 Total IO (KB), bandwidth (KB/sec), IOPS, runtime (msec)
2110 Submission latency: min, max, mean, stdev (usec)
2111 Completion latency: min, max, mean, stdev(usec)
2112 Completion latency percentiles: 20 fields (see below)
2113 Total latency: min, max, mean, stdev (usec)
2114 Bw (KB/s): min, max, aggregate percentage of total, mean, stdev
2115 CPU usage: user, system, context switches, major faults, minor faults
2116 IO depths: <=1, 2, 4, 8, 16, 32, >=64
2117 IO latencies microseconds: <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000
2118 IO latencies milliseconds: <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000, 2000, >=2000
2119 Disk utilization: Disk name, Read ios, write ios,
2120 Read merges, write merges,
2121 Read ticks, write ticks,
2122 Time spent in queue, disk utilization percentage
2123 Additional Info (dependent on continue_on_error, default off): total # errors, first error code
2125 Additional Info (dependent on description being set): Text description
2127 Completion latency percentiles can be a grouping of up to 20 sets, so
2128 for the terse output fio writes all of them. Each field will look like this:
2132 which is the Xth percentile, and the usec latency associated with it.
2134 For disk utilization, all disks used by fio are shown. So for each disk
2135 there will be a disk utilization section.
2138 8.0 Trace file format
2139 ---------------------
2140 There are two trace file format that you can encounter. The older (v1) format
2141 is unsupported since version 1.20-rc3 (March 2008). It will still be described
2142 below in case that you get an old trace and want to understand it.
2144 In any case the trace is a simple text file with a single action per line.
2147 8.1 Trace file format v1
2148 ------------------------
2149 Each line represents a single io action in the following format:
2153 where rw=0/1 for read/write, and the offset and length entries being in bytes.
2155 This format is not supported in Fio versions => 1.20-rc3.
2158 8.2 Trace file format v2
2159 ------------------------
2160 The second version of the trace file format was added in Fio version 1.17.
2161 It allows to access more then one file per trace and has a bigger set of
2162 possible file actions.
2164 The first line of the trace file has to be:
2168 Following this can be lines in two different formats, which are described below.
2170 The file management format:
2174 The filename is given as an absolute path. The action can be one of these:
2176 add Add the given filename to the trace
2177 open Open the file with the given filename. The filename has to have
2178 been added with the add action before.
2179 close Close the file with the given filename. The file has to have been
2183 The file io action format:
2185 filename action offset length
2187 The filename is given as an absolute path, and has to have been added and opened
2188 before it can be used with this format. The offset and length are given in
2189 bytes. The action can be one of these:
2191 wait Wait for 'offset' microseconds. Everything below 100 is discarded.
2192 The time is relative to the previous wait statement.
2193 read Read 'length' bytes beginning from 'offset'
2194 write Write 'length' bytes beginning from 'offset'
2195 sync fsync() the file
2196 datasync fdatasync() the file
2197 trim trim the given file from the given 'offset' for 'length' bytes
2200 9.0 CPU idleness profiling
2201 --------------------------
2202 In some cases, we want to understand CPU overhead in a test. For example,
2203 we test patches for the specific goodness of whether they reduce CPU usage.
2204 fio implements a balloon approach to create a thread per CPU that runs at
2205 idle priority, meaning that it only runs when nobody else needs the cpu.
2206 By measuring the amount of work completed by the thread, idleness of each
2207 CPU can be derived accordingly.
2209 An unit work is defined as touching a full page of unsigned characters. Mean
2210 and standard deviation of time to complete an unit work is reported in "unit
2211 work" section. Options can be chosen to report detailed percpu idleness or
2212 overall system idleness by aggregating percpu stats.
2215 10.0 Verification and triggers
2216 ------------------------------
2217 Fio is usually run in one of two ways, when data verification is done. The
2218 first is a normal write job of some sort with verify enabled. When the
2219 write phase has completed, fio switches to reads and verifies everything
2220 it wrote. The second model is running just the write phase, and then later
2221 on running the same job (but with reads instead of writes) to repeat the
2222 same IO patterns and verify the contents. Both of these methods depend
2223 on the write phase being completed, as fio otherwise has no idea how much
2226 With verification triggers, fio supports dumping the current write state
2227 to local files. Then a subsequent read verify workload can load this state
2228 and know exactly where to stop. This is useful for testing cases where
2229 power is cut to a server in a managed fashion, for instance.
2231 A verification trigger consists of two things:
2233 1) Storing the write state of each job
2234 2) Executing a trigger command
2236 The write state is relatively small, on the order of hundreds of bytes
2237 to single kilobytes. It contains information on the number of completions
2238 done, the last X completions, etc.
2240 A trigger is invoked either through creation ('touch') of a specified
2241 file in the system, or through a timeout setting. If fio is run with
2242 --trigger-file=/tmp/trigger-file, then it will continually check for
2243 the existence of /tmp/trigger-file. When it sees this file, it will
2244 fire off the trigger (thus saving state, and executing the trigger
2247 For client/server runs, there's both a local and remote trigger. If
2248 fio is running as a server backend, it will send the job states back
2249 to the client for safe storage, then execute the remote trigger, if
2250 specified. If a local trigger is specified, the server will still send
2251 back the write state, but the client will then execute the trigger.
2253 10.1 Verification trigger example
2254 ---------------------------------
2255 Lets say we want to run a powercut test on the remote machine 'server'.
2256 Our write workload is in write-test.fio. We want to cut power to 'server'
2257 at some point during the run, and we'll run this test from the safety
2258 or our local machine, 'localbox'. On the server, we'll start the fio
2261 server# fio --server
2263 and on the client, we'll fire off the workload:
2265 localbox$ fio --client=server --trigger-file=/tmp/my-trigger --trigger-remote="bash -c \"echo b > /proc/sysrq-triger\""
2267 We set /tmp/my-trigger as the trigger file, and we tell fio to execute
2269 echo b > /proc/sysrq-trigger
2271 on the server once it has received the trigger and sent us the write
2272 state. This will work, but it's not _really_ cutting power to the server,
2273 it's merely abruptly rebooting it. If we have a remote way of cutting
2274 power to the server through IPMI or similar, we could do that through
2275 a local trigger command instead. Lets assume we have a script that does
2276 IPMI reboot of a given hostname, ipmi-reboot. On localbox, we could
2277 then have run fio with a local trigger instead:
2279 localbox$ fio --client=server --trigger-file=/tmp/my-trigger --trigger="ipmi-reboot server"
2281 For this case, fio would wait for the server to send us the write state,
2282 then execute 'ipmi-reboot server' when that happened.
2284 10.2 Loading verify state
2285 -------------------------
2286 To load store write state, read verification job file must contain
2287 the verify_state_load option. If that is set, fio will load the previously
2288 stored state. For a local fio run this is done by loading the files directly,
2289 and on a client/server run, the server backend will ask the client to send
2290 the files over and load them from there.
2293 11.0 Log File Formats
2294 ---------------------
2296 Fio supports a variety of log file formats, for logging latencies, bandwidth,
2297 and IOPS. The logs share a common format, which looks like this:
2299 time (msec), value, data direction, offset
2301 Time for the log entry is always in milliseconds. The value logged depends
2302 on the type of log, it will be one of the following:
2304 Latency log Value is latency in usecs
2305 Bandwidth log Value is in KB/sec
2306 IOPS log Value is IOPS
2308 Data direction is one of the following:
2314 The offset is the offset, in bytes, from the start of the file, for that
2315 particular IO. The logging of the offset can be toggled with 'log_offset'.
2317 If windowed logging is enabled though 'log_avg_msec', then fio doesn't log
2318 individual IOs. Instead of logs the average values over the specified
2319 period of time. Since 'data direction' and 'offset' are per-IO values,
2320 they aren't applicable if windowed logging is enabled. If windowed logging
2321 is enabled and 'log_max_value' is set, then fio logs maximum values in
2322 that window instead of averages.