Add a suite of NUMA performance benchmarks.
The goal was simulate the behavior and access patterns of real NUMA
workloads, via a wide range of parameters, so this tool goes well
beyond simple bzero() measurements that most NUMA micro-benchmarks use:
- It processes the data and creates a chain of data dependencies,
like a real workload would. Neither the compiler, nor the
kernel (via KSM and other optimizations) nor the CPU can
eliminate parts of the workload.
- It randomizes the initial state and also randomizes the target
addresses of the processing - it's not a simple forward scan
of addresses.
- It provides flexible options to set process, thread and memory
relationship information: -G sets "global" memory shared between
all test processes, -P sets "process" memory shared by all
threads of a process and -T sets "thread" private memory.
- There's a NUMA convergence monitoring and convergence latency
measurement option via -c and -m.
- Micro-sleeps and synchronization can be injected to provoke lock
contention and scheduling, via the -u and -S options. This simulates
IO and contention.
- The -x option instructs the workload to 'perturb' itself artificially
every N seconds, by moving to the first and last CPU of the system
periodically. This way the stability of convergence equilibrium and
the number of steps taken for the scheduler to reach equilibrium again
can be measured.
- The amount of work can be specified via the -l loop count, and/or
via a -s seconds-timeout value.
- CPU and node memory binding options, to test hard binding scenarios.
THP can be turned on and off via madvise() calls.
- Live reporting of convergence progress in an 'at glance' output format.
Printing of convergence and deconvergence events.
The 'perf bench numa mem -a' option will start an array of about 30
individual tests that will each output such measurements:
# Running 5x5-bw-thread, "perf bench numa mem -p 5 -t 5 -P 512 -s 20 -zZ0q --thp 1"
5x5-bw-thread, 20.276, secs, runtime-max/thread
5x5-bw-thread, 20.004, secs, runtime-min/thread
5x5-bw-thread, 20.155, secs, runtime-avg/thread
5x5-bw-thread, 0.671, %, spread-runtime/thread
5x5-bw-thread, 21.153, GB, data/thread
5x5-bw-thread, 528.818, GB, data-total
5x5-bw-thread, 0.959, nsecs, runtime/byte/thread
5x5-bw-thread, 1.043, GB/sec, thread-speed
5x5-bw-thread, 26.081, GB/sec, total-speed
See the help text and the code for more details.
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Hugh Dickins <hughd@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
perf defines both __used and __unused variables to use for marking
unused variables. The variable __used is defined to
__attribute__((__unused__)), which contradicts the kernel definition to
__attribute__((__used__)) for new gcc versions. On Android, __used is
also defined in system headers and this leads to warnings like: warning:
'__used__' attribute ignored
__unused is not defined in the kernel and is not a standard definition.
If __unused is included everywhere instead of __used, this leads to
conflicts with glibc headers, since glibc has a variables with this name
in its headers.
The best approach is to use __maybe_unused, the definition used in the
kernel for __attribute__((unused)). In this way there is only one
definition in perf sources (instead of 2 definitions that point to the
same thing: __used and __unused) and it works on both Linux and Android.
This patch simply replaces all instances of __used and __unused with
__maybe_unused.
Signed-off-by: Irina Tirdea <irina.tirdea@intel.com>
Acked-by: Pekka Enberg <penberg@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/1347315303-29906-7-git-send-email-irina.tirdea@intel.com
[ committer note: fixed up conflict with a116e05 in builtin-sched.c ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
When NDEBUG is defined, the assert macro will be expanded to nothing.
Some assert calls used in perf are also including some functionality
(e.g. system calls), not only validity checks. Therefore, if NDEBUG is
defined, this functionality will be removed along with the assert. Perf
also defines BUG_ON based on assert, so it has the same problem.
Define BUG_ON so that the condition will be executed when NDEBUG is
defined. Replace the assert statements that have these side effects
with BUG_ON.
For defining BUG_ON, use "if (cond) {}" insted of "if (cond) ;" because
in the latter case build fails with "error: suggest braces around empty
body in an ‘if’ statement [-Werror=empty-body]"
Suggested-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Irina Tirdea <irina.tirdea@intel.com>
Reviewed-by: Namhyung Kim <namhyung@kernel.org>
Reviewed-by: Pekka Enberg <penberg@kernel.org>
Cc: David Ahern <dsahern@gmail.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Namhyung Kim <namhyung.kim@lge.com>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Steven Rostedt <rostedt@goodmis.org>
Link: http://lkml.kernel.org/r/1347082551-2394-1-git-send-email-irina.tirdea@intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
As Namhyung Kim pointed, there are confused namings and descriptions of words
"cycle" and "clock" in mem-memset.c and mem-memcpy.c.
With the option "-c" (or "--clock", now renamed as "--cycle"), mem subsystem
measures cost of memset() and memcpy() with cpu-cycles event.
But current mem subsystem source code contains lots of confused variable
namings and descriptions with "clock" (e.g. the variable use_clock). This is a
very bad style because there is another software event named "cpu-clock". This
patch replaces wrong usage of "clock" to "cycle".
v2: modified Documentation/perf-bench.txt for the descriptions of
--cycle option
Signed-off-by: Hitoshi Mitake <h.mitake@gmail.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/1341236777-18457-1-git-send-email-h.mitake@gmail.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
So that we can get the perf bench exec stack fixes and then apply the
remaining fix for the files added after what is in perf/urgent.
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Intended to be able to support the current selection of the preferred
memcpy() implementation, this patch adds the ability to also measure the
two alternative implementations, again by way of using some
pre-processsor replacement.
While on my Westmere system this proves that the movsb based variant is
worse than the movsq based one (since the ERMS feature isn't there), it
also shows that here for the default as well as small sizes the unrolled
variant outperforms the movsq one.
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/4F16D728020000780006D732@nat28.tlf.novell.com
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Since arch/x86/lib/memcpy_64.S implements not only __memcpy, but also
memcpy, without further precautions this function will get chose by the
static linker for resolving all references, and hence the "default"
measurement didn't really measure anything else than the
"x86-64-unrolled" one.
Fix this by renaming (through the pre-processor) the conflicting symbol.
On my Westmere system, the glibc variant turns out to require about 4%
less instructions, but 15% more cycles for the default 1Mb block size
measured.
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/4F16D6FD020000780006D72F@nat28.tlf.novell.com
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
GCC 4.6.0 in Fedora rawhide turned up some compile errors in tools/perf
due to the -Werror=unused-but-set-variable flag.
I've gone through and annotated some of the assignments that had side
effects (ie: return value from a function) with the __used annotation,
and in some cases, just removed unused code.
In a few cases, we were assigning something useful, but not using it in
later parts of the function.
kyle@dreadnought:~/src% gcc --version
gcc (GCC) 4.6.0 20110122 (Red Hat 4.6.0-0.3)
Cc: Ingo Molnar <mingo@redhat.com>
LKML-Reference: <20110124161304.GK27353@bombadil.infradead.org>
Signed-off-by: Kyle McMartin <kyle@redhat.com>
[ committer note: Fixed up the annotation fixes, as that code moved recently ]
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To avoid problems like the one fixed by Stephane Eranian in 3de29ca, now
we'll got this instead:
bench/sched-messaging.c:259: error: negative width in bit-field ‘<anonymous>’
bench/sched-messaging.c:261: error: negative width in bit-field ‘<anonymous>’
Which is rather cryptic, but is how BUILD_BUG_ON_ZERO works, so kernel
hackers should be already used to this.
With it in place found some problems, fixed by changing the affected
variables to sensible types or changed some OPT_INTEGER to OPT_UINTEGER.
Next csets will go thru converting each of the remaining OPT_ so that
review can be made easier by grouping changes per type per patch.
Cc: Frédéric Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Stephane Eranian <eranian@google.com>
Cc: Tom Zanussi <tzanussi@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch adds a new "all" pseudo subsystem and an "all" pseudo
suite. These are for testing all subsystem and its all suite, or
all suite of one subsystem.
(This patch also contains a few trivial comment fixes for
bench/* and output style fixes. I judged that there are no
necessity to make them into individual patch.)
Example of use:
| % ./perf bench sched all # Test all suites of sched subsystem
| # Running sched/messaging benchmark...
| # 20 sender and receiver processes per group
| # 10 groups == 400 processes run
|
| Total time: 0.414 [sec]
|
| # Running sched/pipe benchmark...
| # Extecuted 1000000 pipe operations between two tasks
|
| Total time: 10.999 [sec]
|
| 10.999317 usecs/op
| 90914 ops/sec
|
| % ./perf bench all # Test all suites of all subsystems
| # Running sched/messaging benchmark...
| # 20 sender and receiver processes per group
| # 10 groups == 400 processes run
|
| Total time: 0.420 [sec]
|
| # Running sched/pipe benchmark...
| # Extecuted 1000000 pipe operations between two tasks
|
| Total time: 11.741 [sec]
|
| 11.741346 usecs/op
| 85169 ops/sec
|
| # Running mem/memcpy benchmark...
| # Copying 1MB Bytes from 0x7ff33e920010 to 0x7ff3401ae010 ...
|
| 808.407437 MB/Sec
Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <1260691319-4683-1-git-send-email-mitake@dcl.info.waseda.ac.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
mem-memcpy.c uses perf event system calls to obtain CPU clocks.
And it suddenly dies with BUG_ON() when it running on Linux
doesn't support perf event.
Also fail at calloc() can occur easily when too large
length is passed. Fail of calloc() causes sudden death
with assert().
These behaviours are not friendly. So I fixed the treating of
errors.
Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <1258688237-3797-1-git-send-email-mitake@dcl.info.waseda.ac.jp>
[ v2: improved a few small details ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
'perf bench mem memcpy' is a benchmark suite for measuring memcpy()
performance.
Example on a Intel(R) Core(TM)2 Duo CPU E6850 @ 3.00GHz:
| % perf bench mem memcpy -l 1GB
| # Running mem/memcpy benchmark...
| # Copying 1MB Bytes from 0xb7d98008 to 0xb7e99008 ...
|
| 726.216412 MB/Sec
Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <1258471212-30281-1-git-send-email-mitake@dcl.info.waseda.ac.jp>
[ v2: updated changelog, clarified history of builtin-bench.c ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch modifies builtin-pipe.c for processing common
options. The first option added is "--format".
Users of perf bench will be able to specify output style by
--format.
Usage example:
% ./perf bench sched pipe # with no style specify
(executing 1000000 pipe operations between two tasks)
Total time:5.855 sec
5.855061 usecs/op
170792 ops/sec
% ./perf bench --format=simple sched pipe # specified simple
5.988
Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <1257808802-9420-5-git-send-email-mitake@dcl.info.waseda.ac.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
This patch adds bench/sched-messaging.c.
This benchmark measures performance of scheduler and IPC
mechanisms, and is based on hackbench by Rusty Russell.
Example of usage:
% perf bench sched messaging -g 20 -l 1000 -s
5.432 # in sec
% perf bench sched messaging # run with default
options (20 sender and receiver processes per group)
(10 groups == 400 processes run)
Total time:0.308 sec
% perf bench sched messaging -t -g 20 # # be multi-thread,
with 20 groups (20 sender and receiver threads per group)
(20 groups == 800 threads run)
Total time:0.582 sec
( Rusty is the original author of hackbench.c and he said the code is
and was under the GPLv2 so fine to be merged. )
Signed-off-by: Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
Acked-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: fweisbec@gmail.com
Cc: Jiri Kosina <jkosina@suse.cz>
LKML-Reference: <1257381097-4743-3-git-send-email-mitake@dcl.info.waseda.ac.jp>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
