Re: AIM7 40% regression with 2.6.26-rc1

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Thu, 2008-05-08 at 11:21 +0200, Ingo Molnar wrote:
> * Zhang, Yanmin <yanmin_zhang@linux.intel.com> wrote:
> 
> > >           disk_cp /mnt/shm
> > >   disk_cp (1): cannot open /mnt/shm/tmpa.common
> > >   disk1.c: No such file or directory
> > > 
> > >   [.. etc. a large stream of them .. ]
> > > 
> > > system has 2GB of RAM and tmpfs mounted to the place where aim7 puts its 
> > > work files.
> 
> > My machine has 8GB. To simulate your environment, I reserve 6GB for 
> > hugetlb, then reran the testing and didn't see any failure except: AIM 
> > Multiuser Benchmark - Suite VII Run Beginning
> > 
> > Tasks    jobs/min  jti  jobs/min/task      real       cpu
> >  2000create_shared_memory(): can't create semaphore, pausing...
> > create_shared_memory(): can't create semaphore, pausing...
> 
> that failure message you got worries me - it indicates that your test 
> ran out of IPC semaphores. You can fix it via upping the semaphore 
> limits via:
> 
>    echo "500 32000 128 512" > /proc/sys/kernel/sem
A quick test showed it does work.

Thanks. I need to take shuttle bus or I need walk to home for 2 hours if missing it. :)

> 
> could you check that you still get similar results with this limit 
> fixed?
> 
> note that once i've fixed the semaphore limits it started running fine 
> here. And i see zero idle time during the run on a quad core box.
> 
> here are my numbers:
> 
>   # on v2.6.26-rc1-166-gc0a1811
> 
>   Tasks   Jobs/Min        JTI     Real    CPU     Jobs/sec/task
>   2000    55851.4         93      208.4   793.6   0.4654   # BKL: sleep
>   2000    55402.2         79      210.1   800.1   0.4617
> 
>   2000    55728.4         93      208.9   795.5   0.4644   # BKL: spin
>   2000    55787.2         93      208.7   794.5   0.4649   #
> 
> so the results are the same within noise.
> 
> I'll also check this workload on an 8-way box to make sure it's OK on 
> larger CPU counts too.
> 
> could you double-check your test?
> 
> plus a tty tidbit as well, during the test i saw a few of these:
> 
>  Warning: dev (tty1) tty->count(639) != #fd's(638) in release_dev
>  Warning: dev (tty1) tty->count(462) != #fd's(463) in release_dev
>  Warning: dev (tty1) tty->count(274) != #fd's(275) in release_dev
>  Warning: dev (tty1) tty->count(4) != #fd's(3) in release_dev
>  Warning: dev (tty1) tty->count(164) != #fd's(163) in release_dev
> 
> 	Ingo