8 5. Detailed list of parameters
12 9. CPU idleness profiling
14 1.0 Overview and history
15 ------------------------
16 fio was originally written to save me the hassle of writing special test
17 case programs when I wanted to test a specific workload, either for
18 performance reasons or to find/reproduce a bug. The process of writing
19 such a test app can be tiresome, especially if you have to do it often.
20 Hence I needed a tool that would be able to simulate a given io workload
21 without resorting to writing a tailored test case again and again.
23 A test work load is difficult to define, though. There can be any number
24 of processes or threads involved, and they can each be using their own
25 way of generating io. You could have someone dirtying large amounts of
26 memory in an memory mapped file, or maybe several threads issuing
27 reads using asynchronous io. fio needed to be flexible enough to
28 simulate both of these cases, and many more.
32 The first step in getting fio to simulate a desired io workload, is
33 writing a job file describing that specific setup. A job file may contain
34 any number of threads and/or files - the typical contents of the job file
35 is a global section defining shared parameters, and one or more job
36 sections describing the jobs involved. When run, fio parses this file
37 and sets everything up as described. If we break down a job from top to
38 bottom, it contains the following basic parameters:
40 IO type Defines the io pattern issued to the file(s).
41 We may only be reading sequentially from this
42 file(s), or we may be writing randomly. Or even
43 mixing reads and writes, sequentially or randomly.
45 Block size In how large chunks are we issuing io? This may be
46 a single value, or it may describe a range of
49 IO size How much data are we going to be reading/writing.
51 IO engine How do we issue io? We could be memory mapping the
52 file, we could be using regular read/write, we
53 could be using splice, async io, syslet, or even
56 IO depth If the io engine is async, how large a queuing
57 depth do we want to maintain?
59 IO type Should we be doing buffered io, or direct/raw io?
61 Num files How many files are we spreading the workload over.
63 Num threads How many threads or processes should we spread
66 The above are the basic parameters defined for a workload, in addition
67 there's a multitude of parameters that modify other aspects of how this
73 See the README file for command line parameters, there are only a few
76 Running fio is normally the easiest part - you just give it the job file
77 (or job files) as parameters:
81 and it will start doing what the job_file tells it to do. You can give
82 more than one job file on the command line, fio will serialize the running
83 of those files. Internally that is the same as using the 'stonewall'
84 parameter described in the parameter section.
86 If the job file contains only one job, you may as well just give the
87 parameters on the command line. The command line parameters are identical
88 to the job parameters, with a few extra that control global parameters
89 (see README). For example, for the job file parameter iodepth=2, the
90 mirror command line option would be --iodepth 2 or --iodepth=2. You can
91 also use the command line for giving more than one job entry. For each
92 --name option that fio sees, it will start a new job with that name.
93 Command line entries following a --name entry will apply to that job,
94 until there are no more entries or a new --name entry is seen. This is
95 similar to the job file options, where each option applies to the current
96 job until a new [] job entry is seen.
98 fio does not need to run as root, except if the files or devices specified
99 in the job section requires that. Some other options may also be restricted,
100 such as memory locking, io scheduler switching, and decreasing the nice value.
105 As previously described, fio accepts one or more job files describing
106 what it is supposed to do. The job file format is the classic ini file,
107 where the names enclosed in [] brackets define the job name. You are free
108 to use any ascii name you want, except 'global' which has special meaning.
109 A global section sets defaults for the jobs described in that file. A job
110 may override a global section parameter, and a job file may even have
111 several global sections if so desired. A job is only affected by a global
112 section residing above it. If the first character in a line is a ';' or a
113 '#', the entire line is discarded as a comment.
115 So let's look at a really simple job file that defines two processes, each
116 randomly reading from a 128MB file.
118 ; -- start job file --
129 As you can see, the job file sections themselves are empty as all the
130 described parameters are shared. As no filename= option is given, fio
131 makes up a filename for each of the jobs as it sees fit. On the command
132 line, this job would look as follows:
134 $ fio --name=global --rw=randread --size=128m --name=job1 --name=job2
137 Let's look at an example that has a number of processes writing randomly
140 ; -- start job file --
152 Here we have no global section, as we only have one job defined anyway.
153 We want to use async io here, with a depth of 4 for each file. We also
154 increased the buffer size used to 32KB and define numjobs to 4 to
155 fork 4 identical jobs. The result is 4 processes each randomly writing
156 to their own 64MB file. Instead of using the above job file, you could
157 have given the parameters on the command line. For this case, you would
160 $ fio --name=random-writers --ioengine=libaio --iodepth=4 --rw=randwrite --bs=32k --direct=0 --size=64m --numjobs=4
162 When fio is utilized as a basis of any reasonably large test suite, it might be
163 desirable to share a set of standardized settings across multiple job files.
164 Instead of copy/pasting such settings, any section may pull in an external
165 .fio file with 'include filename' directive, as in the following example:
167 ; -- start job file including.fio --
171 include glob-include.fio
178 include test-include.fio
179 ; -- end job file including.fio --
181 ; -- start job file glob-include.fio --
184 ; -- end job file glob-include.fio --
186 ; -- start job file test-include.fio --
189 ; -- end job file test-include.fio --
191 Settings pulled into a section apply to that section only (except global
192 section). Include directives may be nested in that any included file may
193 contain further include directive(s). Include files may not contain []
197 4.1 Environment variables
198 -------------------------
200 fio also supports environment variable expansion in job files. Any
201 substring of the form "${VARNAME}" as part of an option value (in other
202 words, on the right of the `='), will be expanded to the value of the
203 environment variable called VARNAME. If no such environment variable
204 is defined, or VARNAME is the empty string, the empty string will be
207 As an example, let's look at a sample fio invocation and job file:
209 $ SIZE=64m NUMJOBS=4 fio jobfile.fio
211 ; -- start job file --
218 This will expand to the following equivalent job file at runtime:
220 ; -- start job file --
227 fio ships with a few example job files, you can also look there for
230 4.2 Reserved keywords
231 ---------------------
233 Additionally, fio has a set of reserved keywords that will be replaced
234 internally with the appropriate value. Those keywords are:
236 $pagesize The architecture page size of the running system
237 $mb_memory Megabytes of total memory in the system
238 $ncpus Number of online available CPUs
240 These can be used on the command line or in the job file, and will be
241 automatically substituted with the current system values when the job
242 is run. Simple math is also supported on these keywords, so you can
243 perform actions like:
247 and get that properly expanded to 8 times the size of memory in the
251 5.0 Detailed list of parameters
252 -------------------------------
254 This section describes in details each parameter associated with a job.
255 Some parameters take an option of a given type, such as an integer or
256 a string. Anywhere a numeric value is required, an arithmetic expression
257 may be used, provided it is surrounded by parentheses. Supported operators
267 For time values in expressions, units are microseconds by default. This is
268 different than for time values not in expressions (not enclosed in
269 parentheses). The following types are used:
271 str String. This is a sequence of alpha characters.
272 time Integer with possible time suffix. In seconds unless otherwise
273 specified, use eg 10m for 10 minutes. Accepts s/m/h for seconds,
274 minutes, and hours, and accepts 'ms' (or 'msec') for milliseconds,
275 and 'us' (or 'usec') for microseconds.
276 int SI integer. A whole number value, which may contain a suffix
277 describing the base of the number. Accepted suffixes are k/m/g/t/p,
278 meaning kilo, mega, giga, tera, and peta. The suffix is not case
279 sensitive, and you may also include trailing 'b' (eg 'kb' is the same
280 as 'k'). So if you want to specify 4096, you could either write
281 out '4096' or just give 4k. The suffixes signify base 2 values, so
282 1024 is 1k and 1024k is 1m and so on, unless the suffix is explicitly
283 set to a base 10 value using 'kib', 'mib', 'gib', etc. If that is the
284 case, then 1000 is used as the multiplier. This can be handy for
285 disks, since manufacturers generally use base 10 values when listing
286 the capacity of a drive. If the option accepts an upper and lower
287 range, use a colon ':' or minus '-' to separate such values. May also
288 include a prefix to indicate numbers base. If 0x is used, the number
289 is assumed to be hexadecimal. See irange.
290 bool Boolean. Usually parsed as an integer, however only defined for
291 true and false (1 and 0).
292 irange Integer range with suffix. Allows value range to be given, such
293 as 1024-4096. A colon may also be used as the separator, eg
294 1k:4k. If the option allows two sets of ranges, they can be
295 specified with a ',' or '/' delimiter: 1k-4k/8k-32k. Also see
297 float_list A list of floating numbers, separated by a ':' character.
299 With the above in mind, here follows the complete list of fio job
302 name=str ASCII name of the job. This may be used to override the
303 name printed by fio for this job. Otherwise the job
304 name is used. On the command line this parameter has the
305 special purpose of also signaling the start of a new
308 description=str Text description of the job. Doesn't do anything except
309 dump this text description when this job is run. It's
312 directory=str Prefix filenames with this directory. Used to place files
313 in a different location than "./". See the 'filename' option
314 for escaping certain characters.
316 filename=str Fio normally makes up a filename based on the job name,
317 thread number, and file number. If you want to share
318 files between threads in a job or several jobs, specify
319 a filename for each of them to override the default. If
320 the ioengine used is 'net', the filename is the host, port,
321 and protocol to use in the format of =host,port,protocol.
322 See ioengine=net for more. If the ioengine is file based, you
323 can specify a number of files by separating the names with a
324 ':' colon. So if you wanted a job to open /dev/sda and /dev/sdb
325 as the two working files, you would use
326 filename=/dev/sda:/dev/sdb. On Windows, disk devices are
327 accessed as \\.\PhysicalDrive0 for the first device,
328 \\.\PhysicalDrive1 for the second etc. Note: Windows and
329 FreeBSD prevent write access to areas of the disk containing
330 in-use data (e.g. filesystems).
331 If the wanted filename does need to include a colon, then
332 escape that with a '\' character. For instance, if the filename
333 is "/dev/dsk/foo@3,0:c", then you would use
334 filename="/dev/dsk/foo@3,0\:c". '-' is a reserved name, meaning
335 stdin or stdout. Which of the two depends on the read/write
339 If sharing multiple files between jobs, it is usually necessary
340 to have fio generate the exact names that you want. By default,
341 fio will name a file based on the default file format
342 specification of jobname.jobnumber.filenumber. With this
343 option, that can be customized. Fio will recognize and replace
344 the following keywords in this string:
347 The name of the worker thread or process.
350 The incremental number of the worker thread or
354 The incremental number of the file for that worker
357 To have dependent jobs share a set of files, this option can
358 be set to have fio generate filenames that are shared between
359 the two. For instance, if testfiles.$filenum is specified,
360 file number 4 for any job will be named testfiles.4. The
361 default of $jobname.$jobnum.$filenum will be used if
362 no other format specifier is given.
364 opendir=str Tell fio to recursively add any file it can find in this
365 directory and down the file system tree.
367 lockfile=str Fio defaults to not locking any files before it does
368 IO to them. If a file or file descriptor is shared, fio
369 can serialize IO to that file to make the end result
370 consistent. This is usual for emulating real workloads that
371 share files. The lock modes are:
373 none No locking. The default.
374 exclusive Only one thread/process may do IO,
375 excluding all others.
376 readwrite Read-write locking on the file. Many
377 readers may access the file at the
378 same time, but writes get exclusive
382 rw=str Type of io pattern. Accepted values are:
384 read Sequential reads
385 write Sequential writes
386 randwrite Random writes
387 randread Random reads
388 rw,readwrite Sequential mixed reads and writes
389 randrw Random mixed reads and writes
391 For the mixed io types, the default is to split them 50/50.
392 For certain types of io the result may still be skewed a bit,
393 since the speed may be different. It is possible to specify
394 a number of IO's to do before getting a new offset, this is
395 done by appending a ':<nr>' to the end of the string given.
396 For a random read, it would look like 'rw=randread:8' for
397 passing in an offset modifier with a value of 8. If the
398 suffix is used with a sequential IO pattern, then the value
399 specified will be added to the generated offset for each IO.
400 For instance, using rw=write:4k will skip 4k for every
401 write. It turns sequential IO into sequential IO with holes.
402 See the 'rw_sequencer' option.
404 rw_sequencer=str If an offset modifier is given by appending a number to
405 the rw=<str> line, then this option controls how that
406 number modifies the IO offset being generated. Accepted
409 sequential Generate sequential offset
410 identical Generate the same offset
412 'sequential' is only useful for random IO, where fio would
413 normally generate a new random offset for every IO. If you
414 append eg 8 to randread, you would get a new random offset for
415 every 8 IO's. The result would be a seek for only every 8
416 IO's, instead of for every IO. Use rw=randread:8 to specify
417 that. As sequential IO is already sequential, setting
418 'sequential' for that would not result in any differences.
419 'identical' behaves in a similar fashion, except it sends
420 the same offset 8 number of times before generating a new
423 kb_base=int The base unit for a kilobyte. The defacto base is 2^10, 1024.
424 Storage manufacturers like to use 10^3 or 1000 as a base
425 ten unit instead, for obvious reasons. Allow values are
426 1024 or 1000, with 1024 being the default.
428 unified_rw_reporting=bool Fio normally reports statistics on a per
429 data direction basis, meaning that read, write, and trim are
430 accounted and reported separately. If this option is set,
431 the fio will sum the results and report them as "mixed"
434 randrepeat=bool For random IO workloads, seed the generator in a predictable
435 way so that results are repeatable across repetitions.
437 randseed=int Seed the random number generators based on this seed value, to
438 be able to control what sequence of output is being generated.
439 If not set, the random sequence depends on the randrepeat
442 fallocate=str Whether pre-allocation is performed when laying down files.
445 none Do not pre-allocate space
446 posix Pre-allocate via posix_fallocate()
447 keep Pre-allocate via fallocate() with
448 FALLOC_FL_KEEP_SIZE set
449 0 Backward-compatible alias for 'none'
450 1 Backward-compatible alias for 'posix'
452 May not be available on all supported platforms. 'keep' is only
453 available on Linux.If using ZFS on Solaris this must be set to
454 'none' because ZFS doesn't support it. Default: 'posix'.
456 fadvise_hint=bool By default, fio will use fadvise() to advise the kernel
457 on what IO patterns it is likely to issue. Sometimes you
458 want to test specific IO patterns without telling the
459 kernel about it, in which case you can disable this option.
460 If set, fio will use POSIX_FADV_SEQUENTIAL for sequential
461 IO and POSIX_FADV_RANDOM for random IO.
463 size=int The total size of file io for this job. Fio will run until
464 this many bytes has been transferred, unless runtime is
465 limited by other options (such as 'runtime', for instance).
466 Unless specific nrfiles and filesize options are given,
467 fio will divide this size between the available files
468 specified by the job. If not set, fio will use the full
469 size of the given files or devices. If the files do not
470 exist, size must be given. It is also possible to give
471 size as a percentage between 1 and 100. If size=20% is
472 given, fio will use 20% of the full size of the given
475 io_limit=int Normally fio operates within the region set by 'size', which
476 means that the 'size' option sets both the region and size of
477 IO to be performed. Sometimes that is not what you want. With
478 this option, it is possible to define just the amount of IO
479 that fio should do. For instance, if 'size' is set to 20G and
480 'io_limit' is set to 5G, fio will perform IO within the first
481 20G but exit when 5G have been done.
483 filesize=int Individual file sizes. May be a range, in which case fio
484 will select sizes for files at random within the given range
485 and limited to 'size' in total (if that is given). If not
486 given, each created file is the same size.
488 file_append=bool Perform IO after the end of the file. Normally fio will
489 operate within the size of a file. If this option is set, then
490 fio will append to the file instead. This has identical
491 behavior to setting offset to the size of a file. This option
492 is ignored on non-regular files.
495 fill_fs=bool Sets size to something really large and waits for ENOSPC (no
496 space left on device) as the terminating condition. Only makes
497 sense with sequential write. For a read workload, the mount
498 point will be filled first then IO started on the result. This
499 option doesn't make sense if operating on a raw device node,
500 since the size of that is already known by the file system.
501 Additionally, writing beyond end-of-device will not return
505 bs=int The block size used for the io units. Defaults to 4k. Values
506 can be given for both read and writes. If a single int is
507 given, it will apply to both. If a second int is specified
508 after a comma, it will apply to writes only. In other words,
509 the format is either bs=read_and_write or bs=read,write,trim.
510 bs=4k,8k will thus use 4k blocks for reads, 8k blocks for
511 writes, and 8k for trims. You can terminate the list with
512 a trailing comma. bs=4k,8k, would use the default value for
513 trims.. If you only wish to set the write size, you
514 can do so by passing an empty read size - bs=,8k will set
515 8k for writes and leave the read default value.
518 ba=int At what boundary to align random IO offsets. Defaults to
519 the same as 'blocksize' the minimum blocksize given.
520 Minimum alignment is typically 512b for using direct IO,
521 though it usually depends on the hardware block size. This
522 option is mutually exclusive with using a random map for
523 files, so it will turn off that option.
525 blocksize_range=irange
526 bsrange=irange Instead of giving a single block size, specify a range
527 and fio will mix the issued io block sizes. The issued
528 io unit will always be a multiple of the minimum value
529 given (also see bs_unaligned). Applies to both reads and
530 writes, however a second range can be given after a comma.
533 bssplit=str Sometimes you want even finer grained control of the
534 block sizes issued, not just an even split between them.
535 This option allows you to weight various block sizes,
536 so that you are able to define a specific amount of
537 block sizes issued. The format for this option is:
539 bssplit=blocksize/percentage:blocksize/percentage
541 for as many block sizes as needed. So if you want to define
542 a workload that has 50% 64k blocks, 10% 4k blocks, and
543 40% 32k blocks, you would write:
545 bssplit=4k/10:64k/50:32k/40
547 Ordering does not matter. If the percentage is left blank,
548 fio will fill in the remaining values evenly. So a bssplit
549 option like this one:
551 bssplit=4k/50:1k/:32k/
553 would have 50% 4k ios, and 25% 1k and 32k ios. The percentages
554 always add up to 100, if bssplit is given a range that adds
555 up to more, it will error out.
557 bssplit also supports giving separate splits to reads and
558 writes. The format is identical to what bs= accepts. You
559 have to separate the read and write parts with a comma. So
560 if you want a workload that has 50% 2k reads and 50% 4k reads,
561 while having 90% 4k writes and 10% 8k writes, you would
564 bssplit=2k/50:4k/50,4k/90:8k/10
567 bs_unaligned If this option is given, any byte size value within bsrange
568 may be used as a block range. This typically wont work with
569 direct IO, as that normally requires sector alignment.
571 bs_is_seq_rand If this option is set, fio will use the normal read,write
572 blocksize settings as sequential,random instead. Any random
573 read or write will use the WRITE blocksize settings, and any
574 sequential read or write will use the READ blocksize setting.
576 zero_buffers If this option is given, fio will init the IO buffers to
577 all zeroes. The default is to fill them with random data.
578 The resulting IO buffers will not be completely zeroed,
579 unless scramble_buffers is also turned off.
581 refill_buffers If this option is given, fio will refill the IO buffers
582 on every submit. The default is to only fill it at init
583 time and reuse that data. Only makes sense if zero_buffers
584 isn't specified, naturally. If data verification is enabled,
585 refill_buffers is also automatically enabled.
587 scramble_buffers=bool If refill_buffers is too costly and the target is
588 using data deduplication, then setting this option will
589 slightly modify the IO buffer contents to defeat normal
590 de-dupe attempts. This is not enough to defeat more clever
591 block compression attempts, but it will stop naive dedupe of
592 blocks. Default: true.
594 buffer_compress_percentage=int If this is set, then fio will attempt to
595 provide IO buffer content (on WRITEs) that compress to
596 the specified level. Fio does this by providing a mix of
597 random data and zeroes. Note that this is per block size
598 unit, for file/disk wide compression level that matches
599 this setting, you'll also want to set refill_buffers.
601 buffer_compress_chunk=int See buffer_compress_percentage. This
602 setting allows fio to manage how big the ranges of random
603 data and zeroed data is. Without this set, fio will
604 provide buffer_compress_percentage of blocksize random
605 data, followed by the remaining zeroed. With this set
606 to some chunk size smaller than the block size, fio can
607 alternate random and zeroed data throughout the IO
610 buffer_pattern=str If set, fio will fill the io buffers with this
611 pattern. If not set, the contents of io buffers is defined by
612 the other options related to buffer contents. The setting can
613 be any pattern of bytes, and can be prefixed with 0x for hex
614 values. It may also be a string, where the string must then
617 dedupe_percentage=int If set, fio will generate this percentage of
618 identical buffers when writing. These buffers will be
619 naturally dedupable. The contents of the buffers depend on
620 what other buffer compression settings have been set. It's
621 possible to have the individual buffers either fully
622 compressible, or not at all. This option only controls the
623 distribution of unique buffers.
625 nrfiles=int Number of files to use for this job. Defaults to 1.
627 openfiles=int Number of files to keep open at the same time. Defaults to
628 the same as nrfiles, can be set smaller to limit the number
631 file_service_type=str Defines how fio decides which file from a job to
632 service next. The following types are defined:
634 random Just choose a file at random.
636 roundrobin Round robin over open files. This
639 sequential Finish one file before moving on to
640 the next. Multiple files can still be
641 open depending on 'openfiles'.
643 The string can have a number appended, indicating how
644 often to switch to a new file. So if option random:4 is
645 given, fio will switch to a new random file after 4 ios
648 ioengine=str Defines how the job issues io to the file. The following
651 sync Basic read(2) or write(2) io. lseek(2) is
652 used to position the io location.
654 psync Basic pread(2) or pwrite(2) io.
656 vsync Basic readv(2) or writev(2) IO.
658 psyncv Basic preadv(2) or pwritev(2) IO.
660 libaio Linux native asynchronous io. Note that Linux
661 may only support queued behaviour with
662 non-buffered IO (set direct=1 or buffered=0).
663 This engine defines engine specific options.
665 posixaio glibc posix asynchronous io.
667 solarisaio Solaris native asynchronous io.
669 windowsaio Windows native asynchronous io.
671 mmap File is memory mapped and data copied
672 to/from using memcpy(3).
674 splice splice(2) is used to transfer the data and
675 vmsplice(2) to transfer data from user
678 syslet-rw Use the syslet system calls to make
679 regular read/write async.
681 sg SCSI generic sg v3 io. May either be
682 synchronous using the SG_IO ioctl, or if
683 the target is an sg character device
684 we use read(2) and write(2) for asynchronous
687 null Doesn't transfer any data, just pretends
688 to. This is mainly used to exercise fio
689 itself and for debugging/testing purposes.
691 net Transfer over the network to given host:port.
692 Depending on the protocol used, the hostname,
693 port, listen and filename options are used to
694 specify what sort of connection to make, while
695 the protocol option determines which protocol
697 This engine defines engine specific options.
699 netsplice Like net, but uses splice/vmsplice to
700 map data and send/receive.
701 This engine defines engine specific options.
703 cpuio Doesn't transfer any data, but burns CPU
704 cycles according to the cpuload= and
705 cpucycle= options. Setting cpuload=85
706 will cause that job to do nothing but burn
707 85% of the CPU. In case of SMP machines,
708 use numjobs=<no_of_cpu> to get desired CPU
709 usage, as the cpuload only loads a single
710 CPU at the desired rate.
712 guasi The GUASI IO engine is the Generic Userspace
713 Asyncronous Syscall Interface approach
716 http://www.xmailserver.org/guasi-lib.html
718 for more info on GUASI.
720 rdma The RDMA I/O engine supports both RDMA
721 memory semantics (RDMA_WRITE/RDMA_READ) and
722 channel semantics (Send/Recv) for the
723 InfiniBand, RoCE and iWARP protocols.
725 falloc IO engine that does regular fallocate to
726 simulate data transfer as fio ioengine.
727 DDIR_READ does fallocate(,mode = keep_size,)
728 DDIR_WRITE does fallocate(,mode = 0)
729 DDIR_TRIM does fallocate(,mode = punch_hole)
731 e4defrag IO engine that does regular EXT4_IOC_MOVE_EXT
732 ioctls to simulate defragment activity in
733 request to DDIR_WRITE event
735 rbd IO engine supporting direct access to Ceph
736 Rados Block Devices (RBD) via librbd without
737 the need to use the kernel rbd driver. This
738 ioengine defines engine specific options.
740 gfapi Using Glusterfs libgfapi sync interface to
741 direct access to Glusterfs volumes without
744 gfapi_async Using Glusterfs libgfapi async interface
745 to direct access to Glusterfs volumes without
746 having to go through FUSE. This ioengine
747 defines engine specific options.
749 libhdfs Read and write through Hadoop (HDFS).
750 The 'filename' option is used to specify host,
751 port of the hdfs name-node to connect. This
752 engine interprets offsets a little
753 differently. In HDFS, files once created
754 cannot be modified. So random writes are not
755 possible. To imitate this, libhdfs engine
756 expects bunch of small files to be created
757 over HDFS, and engine will randomly pick a
758 file out of those files based on the offset
759 generated by fio backend. (see the example
760 job file to create such files, use rw=write
761 option). Please note, you might want to set
762 necessary environment variables to work with
763 hdfs/libhdfs properly.
765 external Prefix to specify loading an external
766 IO engine object file. Append the engine
767 filename, eg ioengine=external:/tmp/foo.o
768 to load ioengine foo.o in /tmp.
770 iodepth=int This defines how many io units to keep in flight against
771 the file. The default is 1 for each file defined in this
772 job, can be overridden with a larger value for higher
773 concurrency. Note that increasing iodepth beyond 1 will not
774 affect synchronous ioengines (except for small degress when
775 verify_async is in use). Even async engines may impose OS
776 restrictions causing the desired depth not to be achieved.
777 This may happen on Linux when using libaio and not setting
778 direct=1, since buffered IO is not async on that OS. Keep an
779 eye on the IO depth distribution in the fio output to verify
780 that the achieved depth is as expected. Default: 1.
782 iodepth_batch_submit=int
783 iodepth_batch=int This defines how many pieces of IO to submit at once.
784 It defaults to 1 which means that we submit each IO
785 as soon as it is available, but can be raised to submit
786 bigger batches of IO at the time.
788 iodepth_batch_complete=int This defines how many pieces of IO to retrieve
789 at once. It defaults to 1 which means that we'll ask
790 for a minimum of 1 IO in the retrieval process from
791 the kernel. The IO retrieval will go on until we
792 hit the limit set by iodepth_low. If this variable is
793 set to 0, then fio will always check for completed
794 events before queuing more IO. This helps reduce
795 IO latency, at the cost of more retrieval system calls.
797 iodepth_low=int The low water mark indicating when to start filling
798 the queue again. Defaults to the same as iodepth, meaning
799 that fio will attempt to keep the queue full at all times.
800 If iodepth is set to eg 16 and iodepth_low is set to 4, then
801 after fio has filled the queue of 16 requests, it will let
802 the depth drain down to 4 before starting to fill it again.
804 direct=bool If value is true, use non-buffered io. This is usually
805 O_DIRECT. Note that ZFS on Solaris doesn't support direct io.
806 On Windows the synchronous ioengines don't support direct io.
808 atomic=bool If value is true, attempt to use atomic direct IO. Atomic
809 writes are guaranteed to be stable once acknowledged by
810 the operating system. Only Linux supports O_ATOMIC right
813 buffered=bool If value is true, use buffered io. This is the opposite
814 of the 'direct' option. Defaults to true.
816 offset=int Start io at the given offset in the file. The data before
817 the given offset will not be touched. This effectively
818 caps the file size at real_size - offset.
820 offset_increment=int If this is provided, then the real offset becomes
821 offset + offset_increment * thread_number, where the thread
822 number is a counter that starts at 0 and is incremented for
823 each sub-job (i.e. when numjobs option is specified). This
824 option is useful if there are several jobs which are intended
825 to operate on a file in parallel disjoint segments, with
826 even spacing between the starting points.
828 number_ios=int Fio will normally perform IOs until it has exhausted the size
829 of the region set by size=, or if it exhaust the allocated
830 time (or hits an error condition). With this setting, the
831 range/size can be set independently of the number of IOs to
832 perform. When fio reaches this number, it will exit normally
833 and report status. Note that this does not extend the amount
834 of IO that will be done, it will only stop fio if this
835 condition is met before other end-of-job criteria.
837 fsync=int If writing to a file, issue a sync of the dirty data
838 for every number of blocks given. For example, if you give
839 32 as a parameter, fio will sync the file for every 32
840 writes issued. If fio is using non-buffered io, we may
841 not sync the file. The exception is the sg io engine, which
842 synchronizes the disk cache anyway.
844 fdatasync=int Like fsync= but uses fdatasync() to only sync data and not
846 In FreeBSD and Windows there is no fdatasync(), this falls back to
849 sync_file_range=str:val Use sync_file_range() for every 'val' number of
850 write operations. Fio will track range of writes that
851 have happened since the last sync_file_range() call. 'str'
852 can currently be one or more of:
854 wait_before SYNC_FILE_RANGE_WAIT_BEFORE
855 write SYNC_FILE_RANGE_WRITE
856 wait_after SYNC_FILE_RANGE_WAIT_AFTER
858 So if you do sync_file_range=wait_before,write:8, fio would
859 use SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE for
860 every 8 writes. Also see the sync_file_range(2) man page.
861 This option is Linux specific.
863 overwrite=bool If true, writes to a file will always overwrite existing
864 data. If the file doesn't already exist, it will be
865 created before the write phase begins. If the file exists
866 and is large enough for the specified write phase, nothing
869 end_fsync=bool If true, fsync file contents when a write stage has completed.
871 fsync_on_close=bool If true, fio will fsync() a dirty file on close.
872 This differs from end_fsync in that it will happen on every
873 file close, not just at the end of the job.
875 rwmixread=int How large a percentage of the mix should be reads.
877 rwmixwrite=int How large a percentage of the mix should be writes. If both
878 rwmixread and rwmixwrite is given and the values do not add
879 up to 100%, the latter of the two will be used to override
880 the first. This may interfere with a given rate setting,
881 if fio is asked to limit reads or writes to a certain rate.
882 If that is the case, then the distribution may be skewed.
884 random_distribution=str:float By default, fio will use a completely uniform
885 random distribution when asked to perform random IO. Sometimes
886 it is useful to skew the distribution in specific ways,
887 ensuring that some parts of the data is more hot than others.
888 fio includes the following distribution models:
890 random Uniform random distribution
891 zipf Zipf distribution
892 pareto Pareto distribution
894 When using a zipf or pareto distribution, an input value
895 is also needed to define the access pattern. For zipf, this
896 is the zipf theta. For pareto, it's the pareto power. Fio
897 includes a test program, genzipf, that can be used visualize
898 what the given input values will yield in terms of hit rates.
899 If you wanted to use zipf with a theta of 1.2, you would use
900 random_distribution=zipf:1.2 as the option. If a non-uniform
901 model is used, fio will disable use of the random map.
903 percentage_random=int For a random workload, set how big a percentage should
904 be random. This defaults to 100%, in which case the workload
905 is fully random. It can be set from anywhere from 0 to 100.
906 Setting it to 0 would make the workload fully sequential. Any
907 setting in between will result in a random mix of sequential
908 and random IO, at the given percentages. It is possible to
909 set different values for reads, writes, and trim. To do so,
910 simply use a comma separated list. See blocksize.
912 norandommap Normally fio will cover every block of the file when doing
913 random IO. If this option is given, fio will just get a
914 new random offset without looking at past io history. This
915 means that some blocks may not be read or written, and that
916 some blocks may be read/written more than once. This option
917 is mutually exclusive with verify= if and only if multiple
918 blocksizes (via bsrange=) are used, since fio only tracks
919 complete rewrites of blocks.
921 softrandommap=bool See norandommap. If fio runs with the random block map
922 enabled and it fails to allocate the map, if this option is
923 set it will continue without a random block map. As coverage
924 will not be as complete as with random maps, this option is
927 random_generator=str Fio supports the following engines for generating
928 IO offsets for random IO:
930 tausworthe Strong 2^88 cycle random number generator
931 lfsr Linear feedback shift register generator
933 Tausworthe is a strong random number generator, but it
934 requires tracking on the side if we want to ensure that
935 blocks are only read or written once. LFSR guarantees
936 that we never generate the same offset twice, and it's
937 also less computationally expensive. It's not a true
938 random generator, however, though for IO purposes it's
939 typically good enough. LFSR only works with single
940 block sizes, not with workloads that use multiple block
941 sizes. If used with such a workload, fio may read or write
942 some blocks multiple times.
944 nice=int Run the job with the given nice value. See man nice(2).
946 prio=int Set the io priority value of this job. Linux limits us to
947 a positive value between 0 and 7, with 0 being the highest.
950 prioclass=int Set the io priority class. See man ionice(1).
952 thinktime=int Stall the job x microseconds after an io has completed before
953 issuing the next. May be used to simulate processing being
954 done by an application. See thinktime_blocks and
958 Only valid if thinktime is set - pretend to spend CPU time
959 doing something with the data received, before falling back
960 to sleeping for the rest of the period specified by
964 Only valid if thinktime is set - control how many blocks
965 to issue, before waiting 'thinktime' usecs. If not set,
966 defaults to 1 which will make fio wait 'thinktime' usecs
967 after every block. This effectively makes any queue depth
968 setting redundant, since no more than 1 IO will be queued
969 before we have to complete it and do our thinktime. In
970 other words, this setting effectively caps the queue depth
971 if the latter is larger.
973 rate=int Cap the bandwidth used by this job. The number is in bytes/sec,
974 the normal suffix rules apply. You can use rate=500k to limit
975 reads and writes to 500k each, or you can specify read and
976 writes separately. Using rate=1m,500k would limit reads to
977 1MB/sec and writes to 500KB/sec. Capping only reads or
978 writes can be done with rate=,500k or rate=500k,. The former
979 will only limit writes (to 500KB/sec), the latter will only
982 ratemin=int Tell fio to do whatever it can to maintain at least this
983 bandwidth. Failing to meet this requirement, will cause
984 the job to exit. The same format as rate is used for
985 read vs write separation.
987 rate_iops=int Cap the bandwidth to this number of IOPS. Basically the same
988 as rate, just specified independently of bandwidth. If the
989 job is given a block size range instead of a fixed value,
990 the smallest block size is used as the metric. The same format
991 as rate is used for read vs write separation.
993 rate_iops_min=int If fio doesn't meet this rate of IO, it will cause
994 the job to exit. The same format as rate is used for read vs
997 latency_target=int If set, fio will attempt to find the max performance
998 point that the given workload will run at while maintaining a
999 latency below this target. The values is given in microseconds.
1000 See latency_window and latency_percentile
1002 latency_window=int Used with latency_target to specify the sample window
1003 that the job is run at varying queue depths to test the
1004 performance. The value is given in microseconds.
1006 latency_percentile=float The percentage of IOs that must fall within the
1007 criteria specified by latency_target and latency_window. If not
1008 set, this defaults to 100.0, meaning that all IOs must be equal
1009 or below to the value set by latency_target.
1011 max_latency=int If set, fio will exit the job if it exceeds this maximum
1012 latency. It will exit with an ETIME error.
1014 ratecycle=int Average bandwidth for 'rate' and 'ratemin' over this number
1017 cpumask=int Set the CPU affinity of this job. The parameter given is a
1018 bitmask of allowed CPU's the job may run on. So if you want
1019 the allowed CPUs to be 1 and 5, you would pass the decimal
1020 value of (1 << 1 | 1 << 5), or 34. See man
1021 sched_setaffinity(2). This may not work on all supported
1022 operating systems or kernel versions. This option doesn't
1023 work well for a higher CPU count than what you can store in
1024 an integer mask, so it can only control cpus 1-32. For
1025 boxes with larger CPU counts, use cpus_allowed.
1027 cpus_allowed=str Controls the same options as cpumask, but it allows a text
1028 setting of the permitted CPUs instead. So to use CPUs 1 and
1029 5, you would specify cpus_allowed=1,5. This options also
1030 allows a range of CPUs. Say you wanted a binding to CPUs
1031 1, 5, and 8-15, you would set cpus_allowed=1,5,8-15.
1033 cpus_allowed_policy=str Set the policy of how fio distributes the CPUs
1034 specified by cpus_allowed or cpumask. Two policies are
1037 shared All jobs will share the CPU set specified.
1038 split Each job will get a unique CPU from the CPU set.
1040 'shared' is the default behaviour, if the option isn't
1041 specified. If split is specified, then fio will will assign
1042 one cpu per job. If not enough CPUs are given for the jobs
1043 listed, then fio will roundrobin the CPUs in the set.
1045 numa_cpu_nodes=str Set this job running on spcified NUMA nodes' CPUs. The
1046 arguments allow comma delimited list of cpu numbers,
1047 A-B ranges, or 'all'. Note, to enable numa options support,
1048 fio must be built on a system with libnuma-dev(el) installed.
1050 numa_mem_policy=str Set this job's memory policy and corresponding NUMA
1051 nodes. Format of the argements:
1053 `mode' is one of the following memory policy:
1054 default, prefer, bind, interleave, local
1055 For `default' and `local' memory policy, no node is
1056 needed to be specified.
1057 For `prefer', only one node is allowed.
1058 For `bind' and `interleave', it allow comma delimited
1059 list of numbers, A-B ranges, or 'all'.
1061 startdelay=time Start this job the specified number of seconds after fio
1062 has started. Only useful if the job file contains several
1063 jobs, and you want to delay starting some jobs to a certain
1066 runtime=time Tell fio to terminate processing after the specified number
1067 of seconds. It can be quite hard to determine for how long
1068 a specified job will run, so this parameter is handy to
1069 cap the total runtime to a given time.
1071 time_based If set, fio will run for the duration of the runtime
1072 specified even if the file(s) are completely read or
1073 written. It will simply loop over the same workload
1074 as many times as the runtime allows.
1076 ramp_time=time If set, fio will run the specified workload for this amount
1077 of time before logging any performance numbers. Useful for
1078 letting performance settle before logging results, thus
1079 minimizing the runtime required for stable results. Note
1080 that the ramp_time is considered lead in time for a job,
1081 thus it will increase the total runtime if a special timeout
1082 or runtime is specified.
1084 invalidate=bool Invalidate the buffer/page cache parts for this file prior
1085 to starting io. Defaults to true.
1087 sync=bool Use sync io for buffered writes. For the majority of the
1088 io engines, this means using O_SYNC.
1091 mem=str Fio can use various types of memory as the io unit buffer.
1092 The allowed values are:
1094 malloc Use memory from malloc(3) as the buffers.
1096 shm Use shared memory as the buffers. Allocated
1099 shmhuge Same as shm, but use huge pages as backing.
1101 mmap Use mmap to allocate buffers. May either be
1102 anonymous memory, or can be file backed if
1103 a filename is given after the option. The
1104 format is mem=mmap:/path/to/file.
1106 mmaphuge Use a memory mapped huge file as the buffer
1107 backing. Append filename after mmaphuge, ala
1108 mem=mmaphuge:/hugetlbfs/file
1110 The area allocated is a function of the maximum allowed
1111 bs size for the job, multiplied by the io depth given. Note
1112 that for shmhuge and mmaphuge to work, the system must have
1113 free huge pages allocated. This can normally be checked
1114 and set by reading/writing /proc/sys/vm/nr_hugepages on a
1115 Linux system. Fio assumes a huge page is 4MB in size. So
1116 to calculate the number of huge pages you need for a given
1117 job file, add up the io depth of all jobs (normally one unless
1118 iodepth= is used) and multiply by the maximum bs set. Then
1119 divide that number by the huge page size. You can see the
1120 size of the huge pages in /proc/meminfo. If no huge pages
1121 are allocated by having a non-zero number in nr_hugepages,
1122 using mmaphuge or shmhuge will fail. Also see hugepage-size.
1124 mmaphuge also needs to have hugetlbfs mounted and the file
1125 location should point there. So if it's mounted in /huge,
1126 you would use mem=mmaphuge:/huge/somefile.
1128 iomem_align=int This indiciates the memory alignment of the IO memory buffers.
1129 Note that the given alignment is applied to the first IO unit
1130 buffer, if using iodepth the alignment of the following buffers
1131 are given by the bs used. In other words, if using a bs that is
1132 a multiple of the page sized in the system, all buffers will
1133 be aligned to this value. If using a bs that is not page
1134 aligned, the alignment of subsequent IO memory buffers is the
1135 sum of the iomem_align and bs used.
1138 Defines the size of a huge page. Must at least be equal
1139 to the system setting, see /proc/meminfo. Defaults to 4MB.
1140 Should probably always be a multiple of megabytes, so using
1141 hugepage-size=Xm is the preferred way to set this to avoid
1142 setting a non-pow-2 bad value.
1144 exitall When one job finishes, terminate the rest. The default is
1145 to wait for each job to finish, sometimes that is not the
1148 bwavgtime=int Average the calculated bandwidth over the given time. Value
1149 is specified in milliseconds.
1151 iopsavgtime=int Average the calculated IOPS over the given time. Value
1152 is specified in milliseconds.
1154 create_serialize=bool If true, serialize the file creating for the jobs.
1155 This may be handy to avoid interleaving of data
1156 files, which may greatly depend on the filesystem
1157 used and even the number of processors in the system.
1159 create_fsync=bool fsync the data file after creation. This is the
1162 create_on_open=bool Don't pre-setup the files for IO, just create open()
1163 when it's time to do IO to that file.
1165 create_only=bool If true, fio will only run the setup phase of the job.
1166 If files need to be laid out or updated on disk, only
1167 that will be done. The actual job contents are not
1170 pre_read=bool If this is given, files will be pre-read into memory before
1171 starting the given IO operation. This will also clear
1172 the 'invalidate' flag, since it is pointless to pre-read
1173 and then drop the cache. This will only work for IO engines
1174 that are seekable, since they allow you to read the same data
1175 multiple times. Thus it will not work on eg network or splice
1178 unlink=bool Unlink the job files when done. Not the default, as repeated
1179 runs of that job would then waste time recreating the file
1180 set again and again.
1182 loops=int Run the specified number of iterations of this job. Used
1183 to repeat the same workload a given number of times. Defaults
1186 verify_only Do not perform specified workload---only verify data still
1187 matches previous invocation of this workload. This option
1188 allows one to check data multiple times at a later date
1189 without overwriting it. This option makes sense only for
1190 workloads that write data, and does not support workloads
1191 with the time_based option set.
1193 do_verify=bool Run the verify phase after a write phase. Only makes sense if
1194 verify is set. Defaults to 1.
1196 verify=str If writing to a file, fio can verify the file contents
1197 after each iteration of the job. The allowed values are:
1199 md5 Use an md5 sum of the data area and store
1200 it in the header of each block.
1202 crc64 Use an experimental crc64 sum of the data
1203 area and store it in the header of each
1206 crc32c Use a crc32c sum of the data area and store
1207 it in the header of each block.
1209 crc32c-intel Use hardware assisted crc32c calcuation
1210 provided on SSE4.2 enabled processors. Falls
1211 back to regular software crc32c, if not
1212 supported by the system.
1214 crc32 Use a crc32 sum of the data area and store
1215 it in the header of each block.
1217 crc16 Use a crc16 sum of the data area and store
1218 it in the header of each block.
1220 crc7 Use a crc7 sum of the data area and store
1221 it in the header of each block.
1223 xxhash Use xxhash as the checksum function. Generally
1224 the fastest software checksum that fio
1227 sha512 Use sha512 as the checksum function.
1229 sha256 Use sha256 as the checksum function.
1231 sha1 Use optimized sha1 as the checksum function.
1233 meta Write extra information about each io
1234 (timestamp, block number etc.). The block
1235 number is verified. The io sequence number is
1236 verified for workloads that write data.
1237 See also verify_pattern.
1239 null Only pretend to verify. Useful for testing
1240 internals with ioengine=null, not for much
1243 This option can be used for repeated burn-in tests of a
1244 system to make sure that the written data is also
1245 correctly read back. If the data direction given is
1246 a read or random read, fio will assume that it should
1247 verify a previously written file. If the data direction
1248 includes any form of write, the verify will be of the
1251 verifysort=bool If set, fio will sort written verify blocks when it deems
1252 it faster to read them back in a sorted manner. This is
1253 often the case when overwriting an existing file, since
1254 the blocks are already laid out in the file system. You
1255 can ignore this option unless doing huge amounts of really
1256 fast IO where the red-black tree sorting CPU time becomes
1259 verify_offset=int Swap the verification header with data somewhere else
1260 in the block before writing. Its swapped back before
1263 verify_interval=int Write the verification header at a finer granularity
1264 than the blocksize. It will be written for chunks the
1265 size of header_interval. blocksize should divide this
1268 verify_pattern=str If set, fio will fill the io buffers with this
1269 pattern. Fio defaults to filling with totally random
1270 bytes, but sometimes it's interesting to fill with a known
1271 pattern for io verification purposes. Depending on the
1272 width of the pattern, fio will fill 1/2/3/4 bytes of the
1273 buffer at the time(it can be either a decimal or a hex number).
1274 The verify_pattern if larger than a 32-bit quantity has to
1275 be a hex number that starts with either "0x" or "0X". Use
1278 verify_fatal=bool Normally fio will keep checking the entire contents
1279 before quitting on a block verification failure. If this
1280 option is set, fio will exit the job on the first observed
1283 verify_dump=bool If set, dump the contents of both the original data
1284 block and the data block we read off disk to files. This
1285 allows later analysis to inspect just what kind of data
1286 corruption occurred. Off by default.
1288 verify_async=int Fio will normally verify IO inline from the submitting
1289 thread. This option takes an integer describing how many
1290 async offload threads to create for IO verification instead,
1291 causing fio to offload the duty of verifying IO contents
1292 to one or more separate threads. If using this offload
1293 option, even sync IO engines can benefit from using an
1294 iodepth setting higher than 1, as it allows them to have
1295 IO in flight while verifies are running.
1297 verify_async_cpus=str Tell fio to set the given CPU affinity on the
1298 async IO verification threads. See cpus_allowed for the
1301 verify_backlog=int Fio will normally verify the written contents of a
1302 job that utilizes verify once that job has completed. In
1303 other words, everything is written then everything is read
1304 back and verified. You may want to verify continually
1305 instead for a variety of reasons. Fio stores the meta data
1306 associated with an IO block in memory, so for large
1307 verify workloads, quite a bit of memory would be used up
1308 holding this meta data. If this option is enabled, fio
1309 will write only N blocks before verifying these blocks.
1311 verify_backlog_batch=int Control how many blocks fio will verify
1312 if verify_backlog is set. If not set, will default to
1313 the value of verify_backlog (meaning the entire queue
1314 is read back and verified). If verify_backlog_batch is
1315 less than verify_backlog then not all blocks will be verified,
1316 if verify_backlog_batch is larger than verify_backlog, some
1317 blocks will be verified more than once.
1319 verify_state_save=bool When a job exits during the write phase of a verify
1320 workload, save its current state. This allows fio to replay
1321 up until that point, if the verify state is loaded for the
1322 verify read phase. The format of the filename is, roughly,
1323 <type>-<jobname>-<jobindex>-verify.state. <type> is "local"
1324 for a local run, "sock" for a client/server socket connection,
1325 and "ip" (192.168.0.1, for instance) for a networked
1326 client/server connection.
1328 verify_state_load=bool If a verify termination trigger was used, fio stores
1329 the current write state of each thread. This can be used at
1330 verification time so that fio knows how far it should verify.
1331 Without this information, fio will run a full verification
1332 pass, according to the settings in the job file used.
1335 wait_for_previous Wait for preceding jobs in the job file to exit, before
1336 starting this one. Can be used to insert serialization
1337 points in the job file. A stone wall also implies starting
1338 a new reporting group.
1340 new_group Start a new reporting group. See: group_reporting.
1342 numjobs=int Create the specified number of clones of this job. May be
1343 used to setup a larger number of threads/processes doing
1344 the same thing. Each thread is reported separately; to see
1345 statistics for all clones as a whole, use group_reporting in
1346 conjunction with new_group.
1348 group_reporting It may sometimes be interesting to display statistics for
1349 groups of jobs as a whole instead of for each individual job.
1350 This is especially true if 'numjobs' is used; looking at
1351 individual thread/process output quickly becomes unwieldy.
1352 To see the final report per-group instead of per-job, use
1353 'group_reporting'. Jobs in a file will be part of the same
1354 reporting group, unless if separated by a stonewall, or by
1357 thread fio defaults to forking jobs, however if this option is
1358 given, fio will use pthread_create(3) to create threads
1361 zonesize=int Divide a file into zones of the specified size. See zoneskip.
1363 zoneskip=int Skip the specified number of bytes when zonesize data has
1364 been read. The two zone options can be used to only do
1365 io on zones of a file.
1367 write_iolog=str Write the issued io patterns to the specified file. See
1368 read_iolog. Specify a separate file for each job, otherwise
1369 the iologs will be interspersed and the file may be corrupt.
1371 read_iolog=str Open an iolog with the specified file name and replay the
1372 io patterns it contains. This can be used to store a
1373 workload and replay it sometime later. The iolog given
1374 may also be a blktrace binary file, which allows fio
1375 to replay a workload captured by blktrace. See blktrace
1376 for how to capture such logging data. For blktrace replay,
1377 the file needs to be turned into a blkparse binary data
1378 file first (blkparse <device> -o /dev/null -d file_for_fio.bin).
1380 replay_no_stall=int When replaying I/O with read_iolog the default behavior
1381 is to attempt to respect the time stamps within the log and
1382 replay them with the appropriate delay between IOPS. By
1383 setting this variable fio will not respect the timestamps and
1384 attempt to replay them as fast as possible while still
1385 respecting ordering. The result is the same I/O pattern to a
1386 given device, but different timings.
1388 replay_redirect=str While replaying I/O patterns using read_iolog the
1389 default behavior is to replay the IOPS onto the major/minor
1390 device that each IOP was recorded from. This is sometimes
1391 undesirable because on a different machine those major/minor
1392 numbers can map to a different device. Changing hardware on
1393 the same system can also result in a different major/minor
1394 mapping. Replay_redirect causes all IOPS to be replayed onto
1395 the single specified device regardless of the device it was
1396 recorded from. i.e. replay_redirect=/dev/sdc would cause all
1397 IO in the blktrace to be replayed onto /dev/sdc. This means
1398 multiple devices will be replayed onto a single, if the trace
1399 contains multiple devices. If you want multiple devices to be
1400 replayed concurrently to multiple redirected devices you must
1401 blkparse your trace into separate traces and replay them with
1402 independent fio invocations. Unfortuantely this also breaks
1403 the strict time ordering between multiple device accesses.
1405 write_bw_log=str If given, write a bandwidth log of the jobs in this job
1406 file. Can be used to store data of the bandwidth of the
1407 jobs in their lifetime. The included fio_generate_plots
1408 script uses gnuplot to turn these text files into nice
1409 graphs. See write_lat_log for behaviour of given
1410 filename. For this option, the suffix is _bw.x.log, where
1411 x is the index of the job (1..N, where N is the number of
1414 write_lat_log=str Same as write_bw_log, except that this option stores io
1415 submission, completion, and total latencies instead. If no
1416 filename is given with this option, the default filename of
1417 "jobname_type.log" is used. Even if the filename is given,
1418 fio will still append the type of log. So if one specifies
1422 The actual log names will be foo_slat.x.log, foo_clat.x.log,
1423 and foo_lat.x.log, where x is the index of the job (1..N,
1424 where N is the number of jobs). This helps fio_generate_plot
1425 fine the logs automatically.
1427 write_iops_log=str Same as write_bw_log, but writes IOPS. If no filename is
1428 given with this option, the default filename of
1429 "jobname_type.x.log" is used,where x is the index of the job
1430 (1..N, where N is the number of jobs). Even if the filename
1431 is given, fio will still append the type of log.
1433 log_avg_msec=int By default, fio will log an entry in the iops, latency,
1434 or bw log for every IO that completes. When writing to the
1435 disk log, that can quickly grow to a very large size. Setting
1436 this option makes fio average the each log entry over the
1437 specified period of time, reducing the resolution of the log.
1440 log_offset=int If this is set, the iolog options will include the byte
1441 offset for the IO entry as well as the other data values.
1443 log_compression=int If this is set, fio will compress the IO logs as
1444 it goes, to keep the memory footprint lower. When a log
1445 reaches the specified size, that chunk is removed and
1446 compressed in the background. Given that IO logs are
1447 fairly highly compressible, this yields a nice memory
1448 savings for longer runs. The downside is that the
1449 compression will consume some background CPU cycles, so
1450 it may impact the run. This, however, is also true if
1451 the logging ends up consuming most of the system memory.
1452 So pick your poison. The IO logs are saved normally at the
1453 end of a run, by decompressing the chunks and storing them
1454 in the specified log file. This feature depends on the
1455 availability of zlib.
1457 log_store_compressed=bool If set, and log_compression is also set,
1458 fio will store the log files in a compressed format. They
1459 can be decompressed with fio, using the --inflate-log
1460 command line parameter. The files will be stored with a
1463 lockmem=int Pin down the specified amount of memory with mlock(2). Can
1464 potentially be used instead of removing memory or booting
1465 with less memory to simulate a smaller amount of memory.
1466 The amount specified is per worker.
1468 exec_prerun=str Before running this job, issue the command specified
1469 through system(3). Output is redirected in a file called
1472 exec_postrun=str After the job completes, issue the command specified
1473 though system(3). Output is redirected in a file called
1474 jobname.postrun.txt.
1476 ioscheduler=str Attempt to switch the device hosting the file to the specified
1477 io scheduler before running.
1479 disk_util=bool Generate disk utilization statistics, if the platform
1480 supports it. Defaults to on.
1482 disable_lat=bool Disable measurements of total latency numbers. Useful
1483 only for cutting back the number of calls to gettimeofday,
1484 as that does impact performance at really high IOPS rates.
1485 Note that to really get rid of a large amount of these
1486 calls, this option must be used with disable_slat and
1489 disable_clat=bool Disable measurements of completion latency numbers. See
1492 disable_slat=bool Disable measurements of submission latency numbers. See
1495 disable_bw=bool Disable measurements of throughput/bandwidth numbers. See
1498 clat_percentiles=bool Enable the reporting of percentiles of
1499 completion latencies.
1501 percentile_list=float_list Overwrite the default list of percentiles
1502 for completion latencies. Each number is a floating
1503 number in the range (0,100], and the maximum length of
1504 the list is 20. Use ':' to separate the numbers, and
1505 list the numbers in ascending order. For example,
1506 --percentile_list=99.5:99.9 will cause fio to report
1507 the values of completion latency below which 99.5% and
1508 99.9% of the observed latencies fell, respectively.
1510 clocksource=str Use the given clocksource as the base of timing. The
1511 supported options are:
1513 gettimeofday gettimeofday(2)
1515 clock_gettime clock_gettime(2)
1517 cpu Internal CPU clock source
1519 cpu is the preferred clocksource if it is reliable, as it
1520 is very fast (and fio is heavy on time calls). Fio will
1521 automatically use this clocksource if it's supported and
1522 considered reliable on the system it is running on, unless
1523 another clocksource is specifically set. For x86/x86-64 CPUs,
1524 this means supporting TSC Invariant.
1526 gtod_reduce=bool Enable all of the gettimeofday() reducing options
1527 (disable_clat, disable_slat, disable_bw) plus reduce
1528 precision of the timeout somewhat to really shrink
1529 the gettimeofday() call count. With this option enabled,
1530 we only do about 0.4% of the gtod() calls we would have
1531 done if all time keeping was enabled.
1533 gtod_cpu=int Sometimes it's cheaper to dedicate a single thread of
1534 execution to just getting the current time. Fio (and
1535 databases, for instance) are very intensive on gettimeofday()
1536 calls. With this option, you can set one CPU aside for
1537 doing nothing but logging current time to a shared memory
1538 location. Then the other threads/processes that run IO
1539 workloads need only copy that segment, instead of entering
1540 the kernel with a gettimeofday() call. The CPU set aside
1541 for doing these time calls will be excluded from other
1542 uses. Fio will manually clear it from the CPU mask of other
1545 continue_on_error=str Normally fio will exit the job on the first observed
1546 failure. If this option is set, fio will continue the job when
1547 there is a 'non-fatal error' (EIO or EILSEQ) until the runtime
1548 is exceeded or the I/O size specified is completed. If this
1549 option is used, there are two more stats that are appended,
1550 the total error count and the first error. The error field
1551 given in the stats is the first error that was hit during the
1554 The allowed values are:
1556 none Exit on any IO or verify errors.
1558 read Continue on read errors, exit on all others.
1560 write Continue on write errors, exit on all others.
1562 io Continue on any IO error, exit on all others.
1564 verify Continue on verify errors, exit on all others.
1566 all Continue on all errors.
1568 0 Backward-compatible alias for 'none'.
1570 1 Backward-compatible alias for 'all'.
1572 ignore_error=str Sometimes you want to ignore some errors during test
1573 in that case you can specify error list for each error type.
1574 ignore_error=READ_ERR_LIST,WRITE_ERR_LIST,VERIFY_ERR_LIST
1575 errors for given error type is separated with ':'. Error
1576 may be symbol ('ENOSPC', 'ENOMEM') or integer.
1578 ignore_error=EAGAIN,ENOSPC:122
1579 This option will ignore EAGAIN from READ, and ENOSPC and
1580 122(EDQUOT) from WRITE.
1582 error_dump=bool If set dump every error even if it is non fatal, true
1583 by default. If disabled only fatal error will be dumped
1585 cgroup=str Add job to this control group. If it doesn't exist, it will
1586 be created. The system must have a mounted cgroup blkio
1587 mount point for this to work. If your system doesn't have it
1588 mounted, you can do so with:
1590 # mount -t cgroup -o blkio none /cgroup
1592 cgroup_weight=int Set the weight of the cgroup to this value. See
1593 the documentation that comes with the kernel, allowed values
1594 are in the range of 100..1000.
1596 cgroup_nodelete=bool Normally fio will delete the cgroups it has created after
1597 the job completion. To override this behavior and to leave
1598 cgroups around after the job completion, set cgroup_nodelete=1.
1599 This can be useful if one wants to inspect various cgroup
1600 files after job completion. Default: false
1602 uid=int Instead of running as the invoking user, set the user ID to
1603 this value before the thread/process does any work.
1605 gid=int Set group ID, see uid.
1607 flow_id=int The ID of the flow. If not specified, it defaults to being a
1608 global flow. See flow.
1610 flow=int Weight in token-based flow control. If this value is used, then
1611 there is a 'flow counter' which is used to regulate the
1612 proportion of activity between two or more jobs. fio attempts
1613 to keep this flow counter near zero. The 'flow' parameter
1614 stands for how much should be added or subtracted to the flow
1615 counter on each iteration of the main I/O loop. That is, if
1616 one job has flow=8 and another job has flow=-1, then there
1617 will be a roughly 1:8 ratio in how much one runs vs the other.
1619 flow_watermark=int The maximum value that the absolute value of the flow
1620 counter is allowed to reach before the job must wait for a
1621 lower value of the counter.
1623 flow_sleep=int The period of time, in microseconds, to wait after the flow
1624 watermark has been exceeded before retrying operations
1626 In addition, there are some parameters which are only valid when a specific
1627 ioengine is in use. These are used identically to normal parameters, with the
1628 caveat that when used on the command line, they must come after the ioengine
1629 that defines them is selected.
1631 [libaio] userspace_reap Normally, with the libaio engine in use, fio will use
1632 the io_getevents system call to reap newly returned events.
1633 With this flag turned on, the AIO ring will be read directly
1634 from user-space to reap events. The reaping mode is only
1635 enabled when polling for a minimum of 0 events (eg when
1636 iodepth_batch_complete=0).
1638 [cpu] cpuload=int Attempt to use the specified percentage of CPU cycles.
1640 [cpu] cpuchunks=int Split the load into cycles of the given time. In
1643 [cpu] exit_on_io_done=bool Detect when IO threads are done, then exit.
1645 [netsplice] hostname=str
1646 [net] hostname=str The host name or IP address to use for TCP or UDP based IO.
1647 If the job is a TCP listener or UDP reader, the hostname is not
1648 used and must be omitted unless it is a valid UDP multicast
1651 [netsplice] port=int
1652 [net] port=int The TCP or UDP port to bind to or connect to. If this is used
1653 with numjobs to spawn multiple instances of the same job type, then this will
1654 be the starting port number since fio will use a range of ports.
1656 [netsplice] interface=str
1657 [net] interface=str The IP address of the network interface used to send or
1658 receive UDP multicast
1661 [net] ttl=int Time-to-live value for outgoing UDP multicast packets.
1664 [netsplice] nodelay=bool
1665 [net] nodelay=bool Set TCP_NODELAY on TCP connections.
1667 [netsplice] protocol=str
1668 [netsplice] proto=str
1670 [net] proto=str The network protocol to use. Accepted values are:
1672 tcp Transmission control protocol
1673 tcpv6 Transmission control protocol V6
1674 udp User datagram protocol
1675 udpv6 User datagram protocol V6
1676 unix UNIX domain socket
1678 When the protocol is TCP or UDP, the port must also be given,
1679 as well as the hostname if the job is a TCP listener or UDP
1680 reader. For unix sockets, the normal filename option should be
1681 used and the port is invalid.
1683 [net] listen For TCP network connections, tell fio to listen for incoming
1684 connections rather than initiating an outgoing connection. The
1685 hostname must be omitted if this option is used.
1687 [net] pingpong Normaly a network writer will just continue writing data, and
1688 a network reader will just consume packages. If pingpong=1
1689 is set, a writer will send its normal payload to the reader,
1690 then wait for the reader to send the same payload back. This
1691 allows fio to measure network latencies. The submission
1692 and completion latencies then measure local time spent
1693 sending or receiving, and the completion latency measures
1694 how long it took for the other end to receive and send back.
1695 For UDP multicast traffic pingpong=1 should only be set for a
1696 single reader when multiple readers are listening to the same
1699 [net] window_size Set the desired socket buffer size for the connection.
1701 [net] mss Set the TCP maximum segment size (TCP_MAXSEG).
1703 [e4defrag] donorname=str
1704 File will be used as a block donor(swap extents between files)
1705 [e4defrag] inplace=int
1706 Configure donor file blocks allocation strategy
1707 0(default): Preallocate donor's file on init
1708 1 : allocate space immidietly inside defragment event,
1709 and free right after event
1713 6.0 Interpreting the output
1714 ---------------------------
1716 fio spits out a lot of output. While running, fio will display the
1717 status of the jobs created. An example of that would be:
1719 Threads: 1: [_r] [24.8% done] [ 13509/ 8334 kb/s] [eta 00h:01m:31s]
1721 The characters inside the square brackets denote the current status of
1722 each thread. The possible values (in typical life cycle order) are:
1726 P Thread setup, but not started.
1728 I Thread initialized, waiting or generating necessary data.
1729 p Thread running pre-reading file(s).
1730 R Running, doing sequential reads.
1731 r Running, doing random reads.
1732 W Running, doing sequential writes.
1733 w Running, doing random writes.
1734 M Running, doing mixed sequential reads/writes.
1735 m Running, doing mixed random reads/writes.
1736 F Running, currently waiting for fsync()
1737 f Running, finishing up (writing IO logs, etc)
1738 V Running, doing verification of written data.
1739 E Thread exited, not reaped by main thread yet.
1741 X Thread reaped, exited with an error.
1742 K Thread reaped, exited due to signal.
1744 Fio will condense the thread string as not to take up more space on the
1745 command line as is needed. For instance, if you have 10 readers and 10
1746 writers running, the output would look like this:
1748 Jobs: 20 (f=20): [R(10),W(10)] [4.0% done] [2103MB/0KB/0KB /s] [538K/0/0 iops] [eta 57m:36s]
1750 Fio will still maintain the ordering, though. So the above means that jobs
1751 1..10 are readers, and 11..20 are writers.
1753 The other values are fairly self explanatory - number of threads
1754 currently running and doing io, rate of io since last check (read speed
1755 listed first, then write speed), and the estimated completion percentage
1756 and time for the running group. It's impossible to estimate runtime of
1757 the following groups (if any). Note that the string is displayed in order,
1758 so it's possible to tell which of the jobs are currently doing what. The
1759 first character is the first job defined in the job file, and so forth.
1761 When fio is done (or interrupted by ctrl-c), it will show the data for
1762 each thread, group of threads, and disks in that order. For each data
1763 direction, the output looks like:
1765 Client1 (g=0): err= 0:
1766 write: io= 32MB, bw= 666KB/s, iops=89 , runt= 50320msec
1767 slat (msec): min= 0, max= 136, avg= 0.03, stdev= 1.92
1768 clat (msec): min= 0, max= 631, avg=48.50, stdev=86.82
1769 bw (KB/s) : min= 0, max= 1196, per=51.00%, avg=664.02, stdev=681.68
1770 cpu : usr=1.49%, sys=0.25%, ctx=7969, majf=0, minf=17
1771 IO depths : 1=0.1%, 2=0.3%, 4=0.5%, 8=99.0%, 16=0.0%, 32=0.0%, >32=0.0%
1772 submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
1773 complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
1774 issued r/w: total=0/32768, short=0/0
1775 lat (msec): 2=1.6%, 4=0.0%, 10=3.2%, 20=12.8%, 50=38.4%, 100=24.8%,
1776 lat (msec): 250=15.2%, 500=0.0%, 750=0.0%, 1000=0.0%, >=2048=0.0%
1778 The client number is printed, along with the group id and error of that
1779 thread. Below is the io statistics, here for writes. In the order listed,
1782 io= Number of megabytes io performed
1783 bw= Average bandwidth rate
1784 iops= Average IOs performed per second
1785 runt= The runtime of that thread
1786 slat= Submission latency (avg being the average, stdev being the
1787 standard deviation). This is the time it took to submit
1788 the io. For sync io, the slat is really the completion
1789 latency, since queue/complete is one operation there. This
1790 value can be in milliseconds or microseconds, fio will choose
1791 the most appropriate base and print that. In the example
1792 above, milliseconds is the best scale. Note: in --minimal mode
1793 latencies are always expressed in microseconds.
1794 clat= Completion latency. Same names as slat, this denotes the
1795 time from submission to completion of the io pieces. For
1796 sync io, clat will usually be equal (or very close) to 0,
1797 as the time from submit to complete is basically just
1798 CPU time (io has already been done, see slat explanation).
1799 bw= Bandwidth. Same names as the xlat stats, but also includes
1800 an approximate percentage of total aggregate bandwidth
1801 this thread received in this group. This last value is
1802 only really useful if the threads in this group are on the
1803 same disk, since they are then competing for disk access.
1804 cpu= CPU usage. User and system time, along with the number
1805 of context switches this thread went through, usage of
1806 system and user time, and finally the number of major
1807 and minor page faults.
1808 IO depths= The distribution of io depths over the job life time. The
1809 numbers are divided into powers of 2, so for example the
1810 16= entries includes depths up to that value but higher
1811 than the previous entry. In other words, it covers the
1812 range from 16 to 31.
1813 IO submit= How many pieces of IO were submitting in a single submit
1814 call. Each entry denotes that amount and below, until
1815 the previous entry - eg, 8=100% mean that we submitted
1816 anywhere in between 5-8 ios per submit call.
1817 IO complete= Like the above submit number, but for completions instead.
1818 IO issued= The number of read/write requests issued, and how many
1820 IO latencies= The distribution of IO completion latencies. This is the
1821 time from when IO leaves fio and when it gets completed.
1822 The numbers follow the same pattern as the IO depths,
1823 meaning that 2=1.6% means that 1.6% of the IO completed
1824 within 2 msecs, 20=12.8% means that 12.8% of the IO
1825 took more than 10 msecs, but less than (or equal to) 20 msecs.
1827 After each client has been listed, the group statistics are printed. They
1828 will look like this:
1830 Run status group 0 (all jobs):
1831 READ: io=64MB, aggrb=22178, minb=11355, maxb=11814, mint=2840msec, maxt=2955msec
1832 WRITE: io=64MB, aggrb=1302, minb=666, maxb=669, mint=50093msec, maxt=50320msec
1834 For each data direction, it prints:
1836 io= Number of megabytes io performed.
1837 aggrb= Aggregate bandwidth of threads in this group.
1838 minb= The minimum average bandwidth a thread saw.
1839 maxb= The maximum average bandwidth a thread saw.
1840 mint= The smallest runtime of the threads in that group.
1841 maxt= The longest runtime of the threads in that group.
1843 And finally, the disk statistics are printed. They will look like this:
1845 Disk stats (read/write):
1846 sda: ios=16398/16511, merge=30/162, ticks=6853/819634, in_queue=826487, util=100.00%
1848 Each value is printed for both reads and writes, with reads first. The
1851 ios= Number of ios performed by all groups.
1852 merge= Number of merges io the io scheduler.
1853 ticks= Number of ticks we kept the disk busy.
1854 io_queue= Total time spent in the disk queue.
1855 util= The disk utilization. A value of 100% means we kept the disk
1856 busy constantly, 50% would be a disk idling half of the time.
1858 It is also possible to get fio to dump the current output while it is
1859 running, without terminating the job. To do that, send fio the USR1 signal.
1860 You can also get regularly timed dumps by using the --status-interval
1861 parameter, or by creating a file in /tmp named fio-dump-status. If fio
1862 sees this file, it will unlink it and dump the current output status.
1868 For scripted usage where you typically want to generate tables or graphs
1869 of the results, fio can output the results in a semicolon separated format.
1870 The format is one long line of values, such as:
1872 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%
1873 A description of this job goes here.
1875 The job description (if provided) follows on a second line.
1877 To enable terse output, use the --minimal command line option. The first
1878 value is the version of the terse output format. If the output has to
1879 be changed for some reason, this number will be incremented by 1 to
1880 signify that change.
1882 Split up, the format is as follows:
1884 terse version, fio version, jobname, groupid, error
1886 Total IO (KB), bandwidth (KB/sec), IOPS, runtime (msec)
1887 Submission latency: min, max, mean, deviation (usec)
1888 Completion latency: min, max, mean, deviation (usec)
1889 Completion latency percentiles: 20 fields (see below)
1890 Total latency: min, max, mean, deviation (usec)
1891 Bw (KB/s): min, max, aggregate percentage of total, mean, deviation
1893 Total IO (KB), bandwidth (KB/sec), IOPS, runtime (msec)
1894 Submission latency: min, max, mean, deviation (usec)
1895 Completion latency: min, max, mean, deviation (usec)
1896 Completion latency percentiles: 20 fields (see below)
1897 Total latency: min, max, mean, deviation (usec)
1898 Bw (KB/s): min, max, aggregate percentage of total, mean, deviation
1899 CPU usage: user, system, context switches, major faults, minor faults
1900 IO depths: <=1, 2, 4, 8, 16, 32, >=64
1901 IO latencies microseconds: <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000
1902 IO latencies milliseconds: <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000, 2000, >=2000
1903 Disk utilization: Disk name, Read ios, write ios,
1904 Read merges, write merges,
1905 Read ticks, write ticks,
1906 Time spent in queue, disk utilization percentage
1907 Additional Info (dependent on continue_on_error, default off): total # errors, first error code
1909 Additional Info (dependent on description being set): Text description
1911 Completion latency percentiles can be a grouping of up to 20 sets, so
1912 for the terse output fio writes all of them. Each field will look like this:
1916 which is the Xth percentile, and the usec latency associated with it.
1918 For disk utilization, all disks used by fio are shown. So for each disk
1919 there will be a disk utilization section.
1922 8.0 Trace file format
1923 ---------------------
1924 There are two trace file format that you can encounter. The older (v1) format
1925 is unsupported since version 1.20-rc3 (March 2008). It will still be described
1926 below in case that you get an old trace and want to understand it.
1928 In any case the trace is a simple text file with a single action per line.
1931 8.1 Trace file format v1
1932 ------------------------
1933 Each line represents a single io action in the following format:
1937 where rw=0/1 for read/write, and the offset and length entries being in bytes.
1939 This format is not supported in Fio versions => 1.20-rc3.
1942 8.2 Trace file format v2
1943 ------------------------
1944 The second version of the trace file format was added in Fio version 1.17.
1945 It allows to access more then one file per trace and has a bigger set of
1946 possible file actions.
1948 The first line of the trace file has to be:
1952 Following this can be lines in two different formats, which are described below.
1954 The file management format:
1958 The filename is given as an absolute path. The action can be one of these:
1960 add Add the given filename to the trace
1961 open Open the file with the given filename. The filename has to have
1962 been added with the add action before.
1963 close Close the file with the given filename. The file has to have been
1967 The file io action format:
1969 filename action offset length
1971 The filename is given as an absolute path, and has to have been added and opened
1972 before it can be used with this format. The offset and length are given in
1973 bytes. The action can be one of these:
1975 wait Wait for 'offset' microseconds. Everything below 100 is discarded.
1976 read Read 'length' bytes beginning from 'offset'
1977 write Write 'length' bytes beginning from 'offset'
1978 sync fsync() the file
1979 datasync fdatasync() the file
1980 trim trim the given file from the given 'offset' for 'length' bytes
1983 9.0 CPU idleness profiling
1984 --------------------------
1985 In some cases, we want to understand CPU overhead in a test. For example,
1986 we test patches for the specific goodness of whether they reduce CPU usage.
1987 fio implements a balloon approach to create a thread per CPU that runs at
1988 idle priority, meaning that it only runs when nobody else needs the cpu.
1989 By measuring the amount of work completed by the thread, idleness of each
1990 CPU can be derived accordingly.
1992 An unit work is defined as touching a full page of unsigned characters. Mean
1993 and standard deviation of time to complete an unit work is reported in "unit
1994 work" section. Options can be chosen to report detailed percpu idleness or
1995 overall system idleness by aggregating percpu stats.