Re: [PATCH 0/3] lockless sys_times and posix_cpu_clock_get

[Andrew Theurer <atheurer@redhat.com>]

> Thanks to the feedback from Oleg, Peter, Mike, and Frederic,
> I seem to have a patch series that manages to do times()
> locklessly, and apparently correctly.


> 
> Oleg points out that the monotonicity alone is not enough of a
> guarantee, but that should probably be attacked separately, since
> that issue is equally present with and without these patches...
> 
> The test case below, slightly changed from the one posted by Spencer
> Candland in 2009, now runs in 11 seconds instead of 5 minutes.
> 
> Is it worthwhile?  There apparently are some real workloads that call
> times() a lot, and I believe Sanjay and Andrew have one sitting around.

Thanks for doing this.  When running a OLTP workload in a KVM VM, we saw a 71% increase in performance!  do_sys_times() was a big bottleneck for us.

-Andrew
> 
> --------
> 
> /*
> 
> Based on the test case from the following bug report, but changed
> to measure utime on a per thread basis. (Rik van Riel)
> 
> https://lkml.org/lkml/2009/11/3/522
> 
> From: Spencer Candland
> Subject: utime/stime decreasing on thread exit
> 
> I am seeing a problem with utime/stime decreasing on thread exit in a
> multi-threaded process.  I have been able to track this regression down
> to the "process wide cpu clocks/timers" changes introduces in
> 2.6.29-rc5, specifically when I revert the following commits I know
> longer see decreasing utime/stime values:
> 
> 4da94d49b2ecb0a26e716a8811c3ecc542c2a65d
> 3fccfd67df79c6351a156eb25a7a514e5f39c4d9
> 7d8e23df69820e6be42bcc41d441f4860e8c76f7
> 4cd4c1b40d40447fb5e7ba80746c6d7ba91d7a53
> 32bd671d6cbeda60dc73be77fa2b9037d9a9bfa0
> 
> I poked around a little, but I am afraid I have to admit that I am not
> familiar enough with how this works to resolve this or suggest a fix.
> 
> I have verified this in happening in kernels 2.6.29-rc5 - 2.6.32-rc6, I
> have been testing this on x86 vanilla kernels, but have also verified it
> on several x86 2.6.29+ distro kernels (fedora and ubuntu).
> 
> I first noticed this on a production environment running Apache with the
> worker MPM, however while tracking this down I put together a simple
> program that has been reliable in showing me utime decreasing, hopefully
> it will be helpful in demonstrating the issue:
> */
> 
> #include <stdio.h>
> #include <pthread.h>
> #include <sys/times.h>
> 
> #define NUM_THREADS 500
> 
> struct tms start;
> 
> void *pound (void *threadid)
> {
>   struct tms end;
>   int oldutime = 0;
>   int utime;
>   int c, i;
>   for (i = 0; i < 10000; i++) {
> 	  for (c = 0; c < 10000; c++);
> 	  times(&end);
> 	  utime = ((int)end.tms_utime - (int)start.tms_utime);
> 	  if (oldutime > utime) {
> 	    printf("utime decreased, was %d, now %d!\n", oldutime, utime);
> 	  }
> 	  oldutime = utime;
>   }
>   pthread_exit(NULL);
> }
> 
> int main()
> {
>   pthread_t th[NUM_THREADS];
>   long i;
>   times(&start);
>   for (i = 0; i < NUM_THREADS; i++) {
>     pthread_create (&th[i], NULL, pound, (void *)i);
>   }
>   pthread_exit(NULL);
>   return 0;
> }
> 
>