microsoft / diskspd
https://github.com/microsoft/diskspd
DISKSPD はストレージ負荷ジェネレーターです。
Windows / Windows Server および Cloud Server インフラストラクチャエンジニアリングチームのパフォーマンステストツール
コマンドラインツールですね。
C:\Users\hoge\Downloads\DiskSpd\amd64>diskspd
Usage
Usage: diskspd [options] target1 [ target2 [ target3 ...] ]
version 2.1.0-dev (2021/7/1)
Valid targets:
file_path
#<physical drive number>
<drive_letter>:
Available options:
-? display usage information
-ag group affinity - affinitize threads round-robin to cores in Processor Groups 0 - n.
Group 0 is filled before Group 1, and so forth.
[default; use -n to disable default affinity]
-ag#,#[,#,...]> advanced CPU affinity - affinitize threads round-robin to the CPUs provided. The g# notation
specifies Processor Groups for the following CPU core #s. Multiple Processor Groups
may be specified, and groups/cores may be repeated. If no group is specified, 0 is assumed.
Additional groups/processors may be added, comma separated, or on separate parameters.
Examples: -a0,1,2 and -ag0,0,1,2 are equivalent.
-ag0,0,1,2,g1,0,1,2 specifies the first three cores in groups 0 and 1.
-ag0,0,1,2 -ag1,0,1,2 is equivalent.
-b<size>[KMGT] block size in bytes or KiB/MiB/GiB/TiB [default=64K]
-B<offs>[KMGTb] base target offset in bytes or KiB/MiB/GiB/TiB/blocks [default=0]
(offset from the beginning of the file)
-c<size>[KMGTb] create files of the given size.
Size can be stated in bytes or KiB/MiB/GiB/TiB/blocks
-C<seconds> cool down time - duration of the test after measurements finished [default=0s].
-D<milliseconds> Capture IOPs statistics in intervals of <milliseconds>; these are per-thread
per-target: text output provides IOPs standard deviation, XML provides the full
IOPs time series in addition. [default=1000, 1 second].
-d<seconds> duration (in seconds) to run test [default=10s]
-f<size>[KMGTb] target size - use only the first <size> bytes or KiB/MiB/GiB/TiB/blocks of the file/disk/partition,
for example to test only the first sectors of a disk
-f<rst> open file with one or more additional access hints
r : the FILE_FLAG_RANDOM_ACCESS hint
s : the FILE_FLAG_SEQUENTIAL_SCAN hint
t : the FILE_ATTRIBUTE_TEMPORARY hint
[default: none]
-F<count> total number of threads (conflicts with -t)
-g<value>[i] throughput per-thread per-target throttled to given value; defaults to bytes per millisecond
With the optional i qualifier the value is IOPS of the specified block size (-b).
Throughput limits cannot be specified when using completion routines (-x)
[default: no limit]
-h deprecated, see -Sh
-i<count> number of IOs per burst; see -j [default: inactive]
-j<milliseconds> interval in <milliseconds> between issuing IO bursts; see -i [default: inactive]
-I<priority> Set IO priority to <priority>. Available values are: 1-very low, 2-low, 3-normal (default)
-l Use large pages for IO buffers
-L measure latency statistics
-n disable default affinity (-a)
-N<vni> specify the flush mode for memory mapped I/O
v : uses the FlushViewOfFile API
n : uses the RtlFlushNonVolatileMemory API
i : uses RtlFlushNonVolatileMemory without waiting for the flush to drain
[default: none]
-o<count> number of outstanding I/O requests per target per thread
(1=synchronous I/O, unless more than 1 thread is specified with -F)
[default=2]
-O<count> number of outstanding I/O requests per thread - for use with -F
(1=synchronous I/O)
-p start parallel sequential I/O operations with the same offset
(ignored if -r is specified, makes sense only with -o2 or greater)
-P<count> enable printing a progress dot after each <count> [default=65536]
completed I/O operations, counted separately by each thread
-r[align[KMGTb]] random I/O aligned to <align> in bytes/KiB/MiB/GiB/TiB/blocks (overrides -s)
[default alignment=block size (-b)]
-rd<dist>[params] specify an non-uniform distribution for random IO in the target
[default uniformly random]
distributions: pct, abs
all: IO% and %Target/Size are cumulative. If the sum of IO% is less than 100% the
remainder is applied to the remainder of the target. An IO% of 0 indicates a gap -
no IO will be issued to that range of the target.
pct : parameter is a combination of IO%/%Target separated by : (colon)
Example: -rdpct90/10:0/10:5/20 specifies 90% of IO in 10% of the target, no IO
next 10%, 5% IO in the next 20% and the remaining 5% of IO in the last 60%
abs : parameter is a combination of IO/Target Size separated by : (colon)
If the actual target size is smaller than the distribution, the relative values of IO%
for the valid elements define the effective distribution.
Example: -rdabs90/10G:0/10G:5/20G specifies 90% of IO in 10GiB of the target, no IO
next 10GiB, 5% IO in the next 20GiB and the remaining 5% of IO in the remaining
capacity of the target. If the target is only 20G, the distribution truncates at
90/10G:0:10G and all IO is directed to the first 10G (equivalent to -f10G).
-rs<percentage> percentage of requests which should be issued randomly. When used, -r may be used to
specify IO alignment (applies to both the random and sequential portions of the load).
Sequential IO runs will be homogeneous if a mixed ratio is specified (-w), and run
lengths will follow a geometric distribution based on the percentage split.
-R[p]<text|xml> output format. With the p prefix, the input profile (command line or XML) is validated and
re-output in the specified format without running load, useful for checking or building
complex profiles.
[default: text]
-s[i][align[KMGTb]] stride size of <align> in bytes/KiB/MiB/GiB/TiB/blocks, alignment/offset between operations
[default=non-interlocked, default alignment=block size (-b)]
By default threads track independent sequential IO offsets starting at offset 0 of the target.
With multiple threads this results in threads overlapping their IOs - see -T to divide
them into multiple separate sequential streams on the target.
With the optional i qualifier (-si) threads interlock on a shared sequential offset.
Interlocked operations may introduce overhead but make it possible to issue a single
sequential stream to a target which responds faster than a one thread can drive.
(ignored if -r specified, -si conflicts with -p, -rs and -T)
-S[bhmruw] control caching behavior [default: caching is enabled, no writethrough]
non-conflicting flags may be combined in any order; ex: -Sbw, -Suw, -Swu
-S equivalent to -Su
-Sb enable caching (default, explicitly stated)
-Sh equivalent -Suw
-Sm enable memory mapped I/O
-Su disable software caching, equivalent to FILE_FLAG_NO_BUFFERING
-Sr disable local caching, with remote sw caching enabled; only valid for remote filesystems
-Sw enable writethrough (no hardware write caching), equivalent to FILE_FLAG_WRITE_THROUGH or
non-temporal writes for memory mapped I/O (-Sm)
-t<count> number of threads per target (conflicts with -F)
-T<offs>[KMGTb] starting stride between I/O operations performed on the same target by different threads
[default=0] (starting offset = base file offset + (thread number * <offs>)
only applies with #threads > 1
-v verbose mode
-w<percentage> percentage of write requests (-w and -w0 are equivalent and result in a read-only workload).
absence of this switch indicates 100% reads
IMPORTANT: a write test will destroy existing data without a warning
-W<seconds> warm up time - duration of the test before measurements start [default=5s]
-x use completion routines instead of I/O Completion Ports
-X<filepath> use an XML file to configure the workload. Combine with -R, -v and -z to override profile defaults.
Targets can be defined in XML profiles as template paths of the form *<integer> (*1, *2, ...).
When run, specify the paths to substitute for the template paths in order on the command line.
The first specified target is *1, second is *2, and so on.
Example: diskspd -Xprof.xml first.bin second.bin (prof.xml using *1 and *2)
-z[seed] set random seed [with no -z, seed=0; with plain -z, seed is based on system run time]
Write buffers:
-Z zero buffers used for write tests
-Zr per IO random buffers used for write tests - this incurrs additional run-time
overhead to create random content and shouln't be compared to results run
without -Zr
-Z<size>[KMGb] use a <size> buffer filled with random data as a source for write operations.
-Z<size>[KMGb],<file> use a <size> buffer filled with data from <file> as a source for write operations.
By default, the write buffers are filled with a repeating pattern (0, 1, 2, ..., 255, 0, 1, ...)
Synchronization:
-ys<eventname> signals event <eventname> before starting the actual run (no warmup)
(creates a notification event if <eventname> does not exist)
-yf<eventname> signals event <eventname> after the actual run finishes (no cooldown)
(creates a notification event if <eventname> does not exist)
-yr<eventname> waits on event <eventname> before starting the run (including warmup)
(creates a notification event if <eventname> does not exist)
-yp<eventname> stops the run when event <eventname> is set; CTRL+C is bound to this event
(creates a notification event if <eventname> does not exist)
-ye<eventname> sets event <eventname> and quits
Event Tracing:
-e<q|c|s> Use query perf timer (qpc), cycle count, or system timer respectively.
[default = q, query perf timer (qpc)]
-ep use paged memory for the NT Kernel Logger [default=non-paged memory]
-ePROCESS process start & end
-eTHREAD thread start & end
-eIMAGE_LOAD image load
-eDISK_IO physical disk IO
-eMEMORY_PAGE_FAULTS all page faults
-eMEMORY_HARD_FAULTS hard faults only
-eNETWORK TCP/IP, UDP/IP send & receive
-eREGISTRY registry calls
Examples:
Create 8192KB file and run read test on it for 1 second:
diskspd -c8192K -d1 testfile.dat
Set block size to 4KB, create 2 threads per file, 32 overlapped (outstanding)
I/O operations per thread, disable all caching mechanisms and run block-aligned random
access read test lasting 10 seconds:
diskspd -b4K -t2 -r -o32 -d10 -Sh testfile.dat
Create two 1GB files, set block size to 4KB, create 2 threads per file, affinitize threads
to CPUs 0 and 1 (each file will have threads affinitized to both CPUs) and run read test
lasting 10 seconds:
diskspd -c1G -b4K -t2 -d10 -a0,1 testfile1.dat testfile2.dat
パラメータが多いですね。
こんなときは、Examples で試してみましょう。
コマンド プロンプト
C:\Users\Hoge\Downloads\DiskSpd\amd64>diskspd -c10g -d20 hogehoge.dat
Command Line: diskspd -c10g -d20 hogehoge.dat
Input parameters:
timespan: 1
-------------
duration: 20s
warm up time: 5s
cool down time: 0s
random seed: 0
path: 'hogehoge.dat'
think time: 0ms
burst size: 0
using software cache
using hardware write cache, writethrough off
performing read test
block size: 64KiB
using sequential I/O (stride: 64KiB)
number of outstanding I/O operations per thread: 2
threads per file: 1
using I/O Completion Ports
IO priority: normal
System information:
computer name: HOGE-PC
start time: 2022/06/20 00:29:58 UTC
Results for timespan 1:
*******************************************************************************
actual test time: 20.00s
thread count: 1
proc count: 8
CPU | Usage | User | Kernel | Idle
-------------------------------------------
0| 71.48%| 10.63%| 60.86%| 28.52%
1| 54.45%| 14.61%| 39.84%| 45.55%
2| 72.89%| 19.69%| 53.20%| 27.11%
3| 66.17%| 18.75%| 47.42%| 33.83%
4| 71.48%| 16.17%| 55.31%| 28.52%
5| 67.66%| 18.52%| 49.14%| 32.34%
6| 83.05%| 18.36%| 64.69%| 16.95%
7| 83.59%| 20.63%| 62.97%| 16.41%
-------------------------------------------
avg.| 71.35%| 17.17%| 54.18%| 28.65%
Total IO
thread | bytes | I/Os | MiB/s | I/O per s | file
------------------------------------------------------------------------------
0 | 9749725184 | 148769 | 464.87 | 7437.86 | hogehoge.dat (10GiB)
------------------------------------------------------------------------------
total: 9749725184 | 148769 | 464.87 | 7437.86
Read IO
thread | bytes | I/Os | MiB/s | I/O per s | file
------------------------------------------------------------------------------
0 | 9749725184 | 148769 | 464.87 | 7437.86 | hogehoge.dat (10GiB)
------------------------------------------------------------------------------
total: 9749725184 | 148769 | 464.87 | 7437.86
Write IO
thread | bytes | I/Os | MiB/s | I/O per s | file
------------------------------------------------------------------------------
0 | 0 | 0 | 0.00 | 0.00 | hogehoge.dat (10GiB)
------------------------------------------------------------------------------
total: 0 | 0 | 0.00 | 0.00
10GB の hogehoge.dat を作成して、20秒間読み込むコマンドです。
C:\Users\Hoge\Downloads\DiskSpd\amd64>diskspd -c10g -d20 hogehoge.dat
タスクマネージャーで見ると負荷は上がっていますね。
C:\Users\Hoge\Downloads\DiskSpd\amd64>dir
ドライブ C のボリューム ラベルは Windows です
ボリューム シリアル番号は E4FE-1FBF です
C:\Users\Hoge\Downloads\DiskSpd\amd64 のディレクトリ
2022/06/20 09:29 <DIR> .
2022/06/20 09:29 <DIR> ..
2021/09/17 13:44 282,872 diskspd.exe
2021/09/16 06:24 3,960,832 diskspd.pdb
2022/06/20 09:33 10,737,418,240 hogehoge.dat
3 個のファイル 10,741,661,944 バイト
2 個のディレクトリ 64,278,204,416 バイトの空き領域
作成したファイル（ここでは、hogehoge.dat）は、残っていますので削除してください。