4 fio is a tool that will spawn a number of threads or processes doing a
5 particular type of io action as specified by the user. fio takes a
6 number of global parameters, each inherited by the thread unless
7 otherwise parameters given to them overriding that setting is given.
8 The typical use of fio is to write a job file matching the io load
15 fio resides in a git repo, the canonical place is:
17 git://git.kernel.dk/fio.git
19 If you are inside a corporate firewall, git:// may not always work for
20 you. In that case you can use the http protocol, path is the same:
22 http://git.kernel.dk/fio.git
24 Snapshots are frequently generated and they include the git meta data as
25 well. You can download them here:
27 http://brick.kernel.dk/snaps/
34 Starting with Debian "Squeeze", fio packages are part of the official
35 Debian repository. http://packages.debian.org/search?keywords=fio
38 Starting with Ubuntu 10.04 LTS (aka "Lucid Lynx"), fio packages are part
39 of the Ubuntu "universe" repository.
40 http://packages.ubuntu.com/search?keywords=fio
43 Pascal Bleser <guru@unixtech.be> has fio RPMs in his repository for SUSE
44 variants, you can find them here:
45 http://linux01.gwdg.de/~pbleser/rpm-navigation.php?cat=System/fio
48 Dag Wieƫrs has RPMs for Red Hat related distros, find them here:
49 http://dag.wieers.com/rpm/packages/fio/
52 Mandriva has integrated fio into their package repository, so installing
53 on that distro should be as easy as typing 'urpmi fio'.
56 Packages for Solaris are available from OpenCSW. Install their pkgutil
57 tool (http://www.opencsw.org/get-it/pkgutil/) and then install fio via
61 Bruce Cran <bruce@cran.org.uk> has fio packages for Windows at
62 http://www.bluestop.org/fio .
68 There's a mailing list associated with fio. It's meant for general
69 discussion, bug reporting, questions, and development - basically anything
70 that has to do with fio. An automated mail detailing recent commits is
71 automatically sent to the list at most daily. The list address is
72 fio@vger.kernel.org, subscribe by sending an email to
73 majordomo@vger.kernel.org with
77 in the body of the email. Archives can be found here:
79 http://www.spinics.net/lists/fio/
81 and archives for the old list can be found here:
83 http://maillist.kernel.dk/fio-devel/
89 Just type 'make' and 'make install'.
91 Note that GNU make is required. On BSD it's available from devel/gmake;
92 on Solaris it's in the SUNWgmake package. On platforms where GNU make
93 isn't the default, type 'gmake' instead of 'make'.
95 If your compile fails with an error like this:
98 In file included from fio.h:23,
100 os/os.h:15:20: error: libaio.h: No such file or directory
101 In file included from gettime.c:8:
102 fio.h:119: error: field 'iocb' has incomplete type
103 make: *** [gettime.o] Error 1
105 Check that you have the libaio development package installed. On RPM
106 based distros, it's typically called libaio-devel.
112 On Windows MinGW (http://www.mingw.org/) is required in order to
113 build fio. To create an MSI installer package install WiX 3.6 from
114 http://wix.sourceforge.net/releases/ and run dobuild.cmd from the
115 os/windows directory.
117 How to compile FIO on Windows 64 bits
119 1. Install Cygwin - search for MinGW and install all MinGW packages.
120 2. Download x86_64-w64-mingw32-gcc-4.7.2-release-win64_rubenvb.7z (http://sourceforge.net/projects/mingw-w64/files) - Select Toolchains targeting Win64 ->
121 Personal Builds -> rubenv -> gcc-4.7-release
122 3. Unzip the directory mingw64 to c:\ (c:\mingw64)
123 4. Add to PATH - c:\mingw64\bin
124 5. Copy c:\mingw64\bin\mingw32-make to c:\mingw64\bin\make.exe
125 6. Download pthreads-20100604.zip (http://sourceforge.net/projects/mingw-w64/files) - Select External binary packages (Win64 hosted) -> pthreads
126 7. Unzip pthreads-20100604.zip
127 8. Unzip pthreads-w64.zip
128 9. Copy pthreadGC2-w64.dll to c:\mingw64\bin
129 10. Copy c:\mingw64\bin\pthreadGC2-w64.dll to c:\mingw64\bin\pthreadGC2.dll
130 11. Open Cygwin Terminal
131 12. Go to fio directory (source files)
141 --debug Enable some debugging options (see below)
142 --output Write output to file
143 --runtime Runtime in seconds
144 --latency-log Generate per-job latency logs
145 --bandwidth-log Generate per-job bandwidth logs
146 --minimal Minimal (terse) output
147 --output-format=type Output format (terse,json,normal)
148 --terse-version=type Terse version output format (default 3, or 2 or 4).
149 --version Print version info and exit
150 --help Print this page
151 --cpuclock-test Perform test/validation of CPU clock
152 --cmdhelp=cmd Print command help, "all" for all of them
153 --enghelp=engine Print ioengine help, or list available ioengines
154 --enghelp=engine,cmd Print help for an ioengine cmd
155 --showcmd Turn a job file into command line options
156 --readonly Turn on safety read-only checks, preventing
158 --eta=when When ETA estimate should be printed
159 May be "always", "never" or "auto"
160 --section=name Only run specified section in job file.
161 Multiple sections can be specified.
162 --alloc-size=kb Set smalloc pool to this size in kb (def 1024)
163 --warnings-fatal Fio parser warnings are fatal
164 --max-jobs Maximum number of threads/processes to support
165 --server=args Start backend server. See Client/Server section.
166 --client=host Connect to specified backend.
167 --idle-prof=option Report cpu idleness on a system or percpu basis
168 (option=system,percpu) or run unit work
169 calibration only (option=calibrate).
172 Any parameters following the options will be assumed to be job files,
173 unless they match a job file parameter. You can add as many as you want,
174 each job file will be regarded as a separate group and fio will stonewall
177 The --readonly switch is an extra safety guard to prevent accidentally
178 turning on a write setting when that is not desired. Fio will only write
179 if rw=write/randwrite/rw/randrw is given, but this extra safety net can
180 be used as an extra precaution. It will also enable a write check in the
181 io engine core to prevent an accidental write due to a fio bug.
183 The debug switch allows adding options that trigger certain logging
184 options in fio. Currently the options are:
186 process Dump info related to processes
187 file Dump info related to file actions
188 io Dump info related to IO queuing
189 mem Dump info related to memory allocations
190 blktrace Dump info related to blktrace setup
191 verify Dump info related to IO verification
192 all Enable all debug options
193 random Dump info related to random offset generation
194 parse Dump info related to option matching and parsing
195 diskutil Dump info related to disk utilization updates
196 job:x Dump info only related to job number x
197 mutex Dump info only related to mutex up/down ops
198 profile Dump info related to profile extensions
199 time Dump info related to internal time keeping
200 ? or help Show available debug options.
202 You can specify as many as you want, eg --debug=file,mem will enable
203 file and memory debugging.
205 The section switch is meant to make it easier to ship a bigger job file
206 instead of several smaller ones. Say you define a job file with light,
207 moderate, and heavy parts. Then you can ask fio to run the given part
208 only by giving it a --section=heavy command line option. The section
209 option only applies to job sections, the reserved 'global' section is
210 always parsed and taken into account.
212 Fio has an internal allocator for shared memory called smalloc. It
213 allocates shared structures from this pool. The pool defaults to 1024k
214 in size, and can grow to 128 pools. If running large jobs with randommap
215 enabled it can run out of memory, in which case the --alloc-size switch
216 is handy for starting with a larger pool size. The backing store is
217 files in /tmp. Fio cleans up after itself, while it is running you
218 may see .fio_smalloc.* files in /tmp.
224 See the HOWTO file for a more detailed description of parameters and what
225 they mean. This file contains the terse version. You can describe big and
226 complex setups with the command line, but generally it's a lot easier to
227 just write a simple job file to describe the workload. The job file format
228 is in the ini style format, as that is easy to read and write for the user.
230 The job file parameters are:
232 name=x Use 'x' as the identifier for this job.
233 description=x 'x' is a text description of the job.
234 directory=x Use 'x' as the top level directory for storing files
235 filename=x Force the use of 'x' as the filename for all files
236 in this thread. If not given, fio will make up
237 a suitable filename based on the thread and file
239 rw=x 'x' may be: read, randread, write, randwrite,
240 rw (read-write mix), randrw (read-write random mix)
241 rwmixcycle=x Base cycle for switching between read and write
243 rwmixread=x 'x' percentage of rw mix ios will be reads. If
244 rwmixwrite is also given, the last of the two will
245 be used if they don't add up to 100%.
246 rwmixwrite=x 'x' percentage of rw mix ios will be writes. See
248 rand_repeatable=x The sequence of random io blocks can be repeatable
249 across runs, if 'x' is 1.
250 size=x Set file size to x bytes (x string can include k/m/g)
251 ioengine=x 'x' may be: aio/libaio/linuxaio for Linux aio,
252 posixaio for POSIX aio, solarisaio for Solaris
253 native async IO, windowsaio for Windows native async IO,
254 sync for regular read/write io,
255 psync for regular pread/pwrite io, vsync for regular
256 readv/writev (with queuing emulation) mmap for mmap'ed
257 io, syslet-rw for syslet driven read/write, splice for
258 using splice/vmsplice, sg for direct SG_IO io, net
259 for network io, rdma for RDMA io, or cpuio for a
260 cycler burner load. sg only works on Linux on
261 SCSI (or SCSI-like devices, such as usb-storage or
262 sata/libata driven) devices. Fio also has a null
263 io engine, which is mainly used for testing
266 iodepth=x For async io, allow 'x' ios in flight
267 overwrite=x If 'x', layout a write file first.
268 nrfiles=x Spread io load over 'x' number of files per job,
270 prio=x Run io at prio X, 0-7 is the kernel allowed range
271 prioclass=x Run io at prio class X
272 bs=x Use 'x' for thread blocksize. May include k/m postfix.
273 bsrange=x-y Mix thread block sizes randomly between x and y. May
274 also include k/m postfix.
275 direct=x 1 for direct IO, 0 for buffered IO
276 thinktime=x "Think" x usec after each io
277 rate=x Throttle rate to x KB/sec
278 ratemin=x Quit if rate of x KB/sec can't be met
279 ratecycle=x ratemin averaged over x msecs
280 cpumask=x Only allow job to run on CPUs defined by mask.
281 cpus_allowed=x Like 'cpumask', but allow text setting of CPU affinity.
282 numa_cpu_nodes=x,y-z Allow job to run on specified NUMA nodes' CPU.
283 numa_mem_policy=m:x,y-z Setup numa memory allocation policy.
284 'm' stands for policy, such as local, interleave,
285 bind, prefer, local. 'x, y-z' are numa node(s) for
286 memory allocation according to policy.
287 fsync=x If writing with buffered IO, fsync after every
288 'x' blocks have been written.
289 end_fsync=x If 'x', run fsync() after end-of-job.
290 startdelay=x Start this thread x seconds after startup
291 runtime=x Terminate x seconds after startup. Can include a
292 normal time suffix if not given in seconds, such as
293 'm' for minutes, 'h' for hours, and 'd' for days.
294 offset=x Start io at offset x (x string can include k/m/g)
295 invalidate=x Invalidate page cache for file prior to doing io
296 sync=x Use sync writes if x and writing buffered IO.
297 mem=x If x == malloc, use malloc for buffers. If x == shm,
298 use shared memory for buffers. If x == mmap, use
300 exitall When one thread quits, terminate the others
301 bwavgtime=x Average bandwidth stats over an x msec window.
302 create_serialize=x If 'x', serialize file creation.
303 create_fsync=x If 'x', run fsync() after file creation.
304 unlink If set, unlink files when done.
305 loops=x Run the job 'x' number of times.
306 verify=x If 'x' == md5, use md5 for verifies. If 'x' == crc32,
307 use crc32 for verifies. md5 is 'safer', but crc32 is
308 a lot faster. Only makes sense for writing to a file.
309 For other types of checksumming, see HOWTO.
310 stonewall Wait for preceeding jobs to end before running.
311 numjobs=x Create 'x' similar entries for this job
312 thread Use pthreads instead of forked jobs
314 zoneskip=y Zone options must be paired. If given, the job
315 will skip y bytes for every x read/written. This
316 can be used to gauge hard drive speed over the entire
317 platter, without reading everything. Both x/y can
318 include k/m/g suffix.
319 read_iolog=x Open and read io pattern from file 'x'. The file format
320 is described in the HOWTO.
321 write_iolog=x Write an iolog to file 'x' in the same format as iolog.
322 The iolog options are exclusive, if both given the
323 read iolog will be performed. Specify a separate file
324 for each job, otherwise the iologs will be interspersed
325 and the file may be corrupt.
326 write_bw_log Write a bandwidth log.
327 write_lat_log Write a latency log.
328 lockmem=x Lock down x amount of memory on the machine, to
329 simulate a machine with less memory available. x can
330 include k/m/g suffix.
331 nice=x Run job at given nice value.
332 exec_prerun=x Run 'x' before job io is begun.
333 exec_postrun=x Run 'x' after job io has finished.
334 ioscheduler=x Use ioscheduler 'x' for this job.
335 cpuload=x For a CPU io thread, percentage of CPU time to attempt
337 cpuchunks=x Split burn cycles into pieces of x usecs.
344 Normally you would run fio as a stand-alone application on the machine
345 where the IO workload should be generated. However, it is also possible to
346 run the frontend and backend of fio separately. This makes it possible to
347 have a fio server running on the machine(s) where the IO workload should
348 be running, while controlling it from another machine.
350 To start the server, you would do:
354 on that machine, where args defines what fio listens to. The arguments
355 are of the form 'type,hostname or IP,port'. 'type' is either 'ip' (or ip4)
356 for TCP/IP v4, 'ip6' for TCP/IP v6, or 'sock' for a local unix domain socket.
357 'hostname' is either a hostname or IP address, and 'port' is the port to
358 listen to (only valid for TCP/IP, not a local socket). Some examples:
362 Start a fio server, listening on all interfaces on the default port (8765).
364 2) fio --server=ip:hostname,4444
366 Start a fio server, listening on IP belonging to hostname and on port 4444.
368 3) fio --server=ip6:::1,4444
370 Start a fio server, listening on IPv6 localhost ::1 and on port 4444.
372 4) fio --server=,4444
374 Start a fio server, listening on all interfaces on port 4444.
376 5) fio --server=1.2.3.4
378 Start a fio server, listening on IP 1.2.3.4 on the default port.
380 6) fio --server=sock:/tmp/fio.sock
382 Start a fio server, listening on the local socket /tmp/fio.sock.
384 When a server is running, you can connect to it from a client. The client
387 fio --local-args --client=server --remote-args <job file(s)>
389 where --local-args are arguments that are local to the client where it is
390 running, 'server' is the connect string, and --remote-args and <job file(s)>
391 are sent to the server. The 'server' string follows the same format as it
392 does on the server side, to allow IP/hostname/socket and port strings.
393 You can connect to multiple clients as well, to do that you could run:
395 fio --client=server2 <job file(s)> --client=server2 <job file(s)>
401 Fio works on (at least) Linux, Solaris, AIX, HP-UX, OSX, NetBSD, Windows
402 and FreeBSD. Some features and/or options may only be available on some of
403 the platforms, typically because those features only apply to that platform
404 (like the solarisaio engine, or the splice engine on Linux).
406 Some features are not available on FreeBSD/Solaris even if they could be
407 implemented, I'd be happy to take patches for that. An example of that is
408 disk utility statistics and (I think) huge page support, support for that
409 does exist in FreeBSD/Solaris.
411 Fio uses pthread mutexes for signalling and locking and FreeBSD does not
412 support process shared pthread mutexes. As a result, only threads are
413 supported on FreeBSD. This could be fixed with sysv ipc locking or
414 other locking alternatives.
416 Other *BSD platforms are untested, but fio should work there almost out
417 of the box. Since I don't do test runs or even compiles on those platforms,
418 your mileage may vary. Sending me patches for other platforms is greatly
419 appreciated. There's a lot of value in having the same test/benchmark tool
420 available on all platforms.
422 Note that POSIX aio is not enabled by default on AIX. If you get messages like:
424 Symbol resolution failed for /usr/lib/libc.a(posix_aio.o) because:
425 Symbol _posix_kaio_rdwr (number 2) is not exported from dependent module /unix.
427 you need to enable POSIX aio. Run the following commands as root:
429 # lsdev -C -l posix_aio0
430 posix_aio0 Defined Posix Asynchronous I/O
431 # cfgmgr -l posix_aio0
432 # lsdev -C -l posix_aio0
433 posix_aio0 Available Posix Asynchronous I/O
435 POSIX aio should work now. To make the change permanent:
437 # chdev -l posix_aio0 -P -a autoconfig='available'
444 Fio was written by Jens Axboe <axboe@kernel.dk> to enable flexible testing
445 of the Linux IO subsystem and schedulers. He got tired of writing
446 specific test applications to simulate a given workload, and found that
447 the existing io benchmark/test tools out there weren't flexible enough
448 to do what he wanted.
450 Jens Axboe <axboe@kernel.dk> 20060905