[clockevents] d5a085970b0: -9.2% will-it-scale.per_thread_ops

[clockevents] d5a085970b0: -9.2% will-it-scale.per_thread_ops

[Fengguang Wu]

[-- Attachment #1: Type: text/plain, Size: 840 bytes --]

Hi Viresh,

I'm not sure how the simple commit will impact will-it-scale/pwrite2
throughput. Just in case you can see more clues.

git://git.linaro.org/people/vireshk/linux tick/ONESHOT-STOPPED
commit d5a085970b0b88b8c82303d2c45f09c9427fd02a ("clockevents: x86: add support for ONESHOT_STOPPED mode")

test case: nhm4/will-it-scale/pwrite2

f4f5df6f142f679  d5a085970b0b88b8c82303d2c  
---------------  -------------------------  
    773859 ~ 0%      -9.2%     702704 ~ 0%  TOTAL will-it-scale.per_thread_ops

Legend:
	~XX%    - stddev percent
	[+-]XX% - change percent

Disclaimer:
Results have been estimated based on internal Intel analysis and are provided
for informational purposes only. Any difference in system hardware or software
design or configuration may affect actual performance.

Thanks,
Fengguang

Re: [clockevents] d5a085970b0: -9.2% will-it-scale.per_thread_ops

[Viresh Kumar]

[-- Attachment #1: Type: text/plain, Size: 449 bytes --]

On 26 June 2014 07:16, Fengguang Wu <fengguang.wu@intel.com> wrote:
> I'm not sure how the simple commit will impact will-it-scale/pwrite2
> throughput. Just in case you can see more clues.
>
> git://git.linaro.org/people/vireshk/linux tick/ONESHOT-STOPPED
> commit d5a085970b0b88b8c82303d2c45f09c9427fd02a ("clockevents: x86: add support for ONESHOT_STOPPED mode")
>
> test case: nhm4/will-it-scale/pwrite2

What is this test all about ?

Re: [clockevents] d5a085970b0: -9.2% will-it-scale.per_thread_ops

[Fengguang Wu]

[-- Attachment #1: Type: text/plain, Size: 682 bytes --]

On Thu, Jun 26, 2014 at 09:35:11AM +0530, Viresh Kumar wrote:
> On 26 June 2014 07:16, Fengguang Wu <fengguang.wu@intel.com> wrote:
> > I'm not sure how the simple commit will impact will-it-scale/pwrite2
> > throughput. Just in case you can see more clues.
> >
> > git://git.linaro.org/people/vireshk/linux tick/ONESHOT-STOPPED
> > commit d5a085970b0b88b8c82303d2c45f09c9427fd02a ("clockevents: x86: add support for ONESHOT_STOPPED mode")
> >
> > test case: nhm4/will-it-scale/pwrite2
> 
> What is this test all about ?

It calls the pwrite2() syscall in parallel, on one single tmpfs file.

https://github.com/antonblanchard/will-it-scale.git

Thanks,
Fengguang

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: pwrite2.c --]
[-- Type: text/x-csrc, Size: 700 bytes --]

#define _XOPEN_SOURCE 500
#include <fcntl.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <assert.h>

#define BUFLEN 4096

static char tmpfile[] = "/tmp/willitscale.XXXXXX";

char *testcase_description = "Same file pwrite";

void testcase_prepare(void)
{
	char buf[BUFLEN];
	int fd = mkstemp(tmpfile);

	memset(buf, 0, sizeof(buf));
	assert(fd >= 0);
	assert(write(fd, buf, sizeof(buf)) == sizeof(buf));
	close(fd);
}

void testcase(unsigned long long *iterations)
{
	char buf[BUFLEN];
	int fd = open(tmpfile, O_RDWR);

	assert(fd >= 0);

	while (1) {
		assert(pwrite(fd, buf, BUFLEN, 0) == BUFLEN);

		(*iterations)++;
	}
}

void testcase_cleanup(void)
{
	unlink(tmpfile);
}

Re: [clockevents] d5a085970b0: -9.2% will-it-scale.per_thread_ops

[Viresh Kumar]

[-- Attachment #1: Type: text/plain, Size: 1467 bytes --]

On 26 June 2014 07:16, Fengguang Wu <fengguang.wu@intel.com> wrote:
> Hi Viresh,
>
> I'm not sure how the simple commit will impact will-it-scale/pwrite2
> throughput. Just in case you can see more clues.
>
> git://git.linaro.org/people/vireshk/linux tick/ONESHOT-STOPPED
> commit d5a085970b0b88b8c82303d2c45f09c9427fd02a ("clockevents: x86: add support for ONESHOT_STOPPED mode")
>
> test case: nhm4/will-it-scale/pwrite2
>
> f4f5df6f142f679  d5a085970b0b88b8c82303d2c
> ---------------  -------------------------
>     773859 ~ 0%      -9.2%     702704 ~ 0%  TOTAL will-it-scale.per_thread_ops
>
> Legend:
>         ~XX%    - stddev percent
>         [+-]XX% - change percent
>
> Disclaimer:
> Results have been estimated based on internal Intel analysis and are provided
> for informational purposes only. Any difference in system hardware or software
> design or configuration may affect actual performance.

Hi Fengguang,

Thanks for your link: https://github.com/antonblanchard/will-it-scale.git

You saw these changes because the simple commit you were talking about
changes behavior of code.

To be specific, we are stopping the clock-event devices now when they
aren't used. i.e. during idle periods or in full dynticks.

I still fail to understand why performance came down by 9.2%..
Adding Frederic in CC, in case he can throw some light. But he is currently
on vacations and I don't expect him to reply immediately :)

Re: [clockevents] d5a085970b0: -9.2% will-it-scale.per_thread_ops

[Dave Hansen]

[-- Attachment #1: Type: text/plain, Size: 357 bytes --]

On 06/25/2014 06:46 PM, Fengguang Wu wrote:
> f4f5df6f142f679  d5a085970b0b88b8c82303d2c  
> ---------------  -------------------------  
>     773859 ~ 0%      -9.2%     702704 ~ 0%  TOTAL will-it-scale.per_thread_ops

Has this impacted more than a single system?  I wonder if that system
was using the HPET where all the others were using the TSC.