Re: [PATCH v2] Perf Bench: Locking Microbenchmark

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Fri, Nov 21, 2014 at 12:57:06PM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Thu, Nov 20, 2014 at 11:06:05AM -0800, Tuan Bui escreveu:
> > Subject: [PATCH] Perf Bench: Locking Microbenchmark
> > 
> > In response to this thread https://lkml.org/lkml/2014/2/11/93, this is
> > a micro benchmark that stresses locking contention in the kernel with
> > creat(2) system call by spawning multiple processes to spam this system
> > call.  This workload generate similar results and contentions in AIM7
> > fserver workload but can generate outputs within seconds.
> > 
> > With the creat system call the contention vary on what locks are used
> > in the particular file system. I have ran this benchmark only on ext4
> > and xfs file system.

I noticed that if control+C it it leaves tons of files in the current
directory, can you please add code to make it handle this? I think that
it would also be better to create a temporary directory, etc.

And please take a look at the edited changelog below, to reflect those
changes on your next attempt to submit this patch, ok? I added an
Example so that people can now at a glance how it changes the existing
output for 'perf bench' and what is the output for 'perf bench locking'.

- Arnaldo

Subject: [PATCH] perf bench: Locking Microbenchmark

In response to this thread https://lkml.org/lkml/2014/2/11/93, this is
a micro benchmark that stresses locking contention in the kernel with
creat(2) system call by spawning multiple processes to spam this system
call.  This workload generate similar results and contentions in AIM7
fserver workload but can generate outputs within seconds.

With the creat system call the contention vary on what locks are used
in the particular file system. I have ran this benchmark only on ext4
and xfs file system.

Running the creat workload on ext4 show contention in the mutex lock
that is used by ext4_orphan_add() and ext4_orphan_del() to add or delete
an inode from the list of inodes. At the same time running the creat
workload on xfs show contention in the spinlock that is used by
xsf_log_commit_cil() to commit a transaction to the Committed Item List.

Here is a comparison of this benchmark with AIM7 running fserver workload
at 500-1000 users along with a perf trace running on ext4 file system.

Test machine is a 8-sockets 80 cores Westmere system HT-off on v3.17-rc6.

	AIM7		AIM7		perf-bench	perf-bench
Users	Jobs/min	Jobs/min/child	Ops/sec		Ops/sec/child
500	119668.25	239.34		104249		208
600	126074.90	210.12		106136		176
700	128662.42	183.80		106175		151
800	119822.05	149.78		106290		132
900	106150.25	117.94		105230		116
1000	104681.29	104.68		106489		106

Perf report for AIM7 fserver:
14.51%	reaim  		[kernel.kallsyms]	[k] osq_lock
4.98%	reaim  		reaim			[.] add_long
4.98%	reaim  		reaim			[.] add_int
4.31%	reaim  		[kernel.kallsyms]	[k] mutex_spin_on_owner
...

Perf report of 'perf bench locking vfs'

22.37%	locking-creat  [kernel.kallsyms]	[k] osq_lock
5.77%	locking-creat  [kernel.kallsyms]	[k] mutex_spin_on_owner
5.31%	locking-creat  [kernel.kallsyms]	[k] _raw_spin_lock
5.15%	locking-creat  [jbd2]			[k] jbd2_journal_put_journal_head
...

Example:

[root@zoo ~]# perf bench
Usage:
	perf bench [<common options>] <collection> <benchmark>
[<options>]

        # List of all available benchmark collections:

         sched: Scheduler and IPC benchmarks
           mem: Memory access benchmarks
          numa: NUMA scheduling and MM benchmarks
         futex: Futex stressing benchmarks
       locking: Kernel locking benchmarks
           all: All benchmarks

[root@zoo ~]# perf bench locking

        # List of available benchmarks for collection 'locking':

           vfs: Benchmark vfs using creat(2)
           all: Run all benchmarks in this suite

[root@zoo ~]# perf bench locking vfs

   100 processes: throughput = 342506 average opts/sec all processes
   100 processes: throughput = 3425 average opts/sec per process

   200 processes: throughput = 341309 average opts/sec all processes
   200 processes: throughput = 1706 average opts/sec per process
<SNIP>

Changes since v1:
- Added -j options to specified jobs per processes.
- Change name of microbenchmark from creat to vfs.
- Change all instances of threads to proccess.