Re: [RFC PATCH 0/4] Gang scheduling in CFS

[Rik van Riel]

On 01/04/2012 09:41 AM, Avi Kivity wrote:
> On 01/04/2012 12:52 PM, Nikunj A Dadhania wrote:
> > On Mon, 02 Jan 2012 11:37:22 +0200, Avi Kivity<avi@xxxxxxxxxx>  wrote:
> > > On 12/31/2011 04:21 AM, Nikunj A Dadhania wrote:
> > > >      GangV2:
> > > >      27.45%       ebizzy  libc-2.12.so            [.] __memcpy_ssse3_back
> > > >      12.12%       ebizzy  [kernel.kallsyms]       [k] clear_page
> > > >       9.22%       ebizzy  [kernel.kallsyms]       [k] __do_page_fault
> > > >       6.91%       ebizzy  [kernel.kallsyms]       [k] flush_tlb_others_ipi
> > > >       4.06%       ebizzy  [kernel.kallsyms]       [k] get_page_from_freelist
> > > >       4.04%       ebizzy  [kernel.kallsyms]       [k] ____pagevec_lru_add
> > > >
> > > >      GangBase:
> > > >      45.08%       ebizzy  [kernel.kallsyms]       [k] flush_tlb_others_ipi
> > > >      15.38%       ebizzy  libc-2.12.so            [.] __memcpy_ssse3_back
> > > >       7.00%       ebizzy  [kernel.kallsyms]       [k] clear_page
> > > >       4.88%       ebizzy  [kernel.kallsyms]       [k] __do_page_fault
> > >
> > > Looping in flush_tlb_others().  Rik, what trace an we run to find out
> > > why PLE directed yield isn't working as expected?
> > I tried some experiments by adding a pause_loop_exits stat in the
> > kvm_vpu_stat.
>
> (that's deprecated, we use tracepoints these days for stats)
>
> > Here are some observation related to Baseline-only(8vm case)
> >
> >                | ple_gap=128 | ple_gap=64 | ple_gap=256 | ple_window=2048
> > --------------+-------------+------------+-------------+----------------
> > EbzyRecords/s |    2247.50  |    2132.75 |    2086.25  |      1835.62
> > PauseExits    | 7928154.00  | 6696342.00 | 7365999.00  |  50319582.00
> >
> > With ple_window = 2048, PauseExits is more than 6times the default case
>
> So it looks like the default is optimal, at least wrt the cases you
> tested and your test workload.

It depends on the workload.

I believe ebizzy synchronously bounces messages around between
userland threads, and may benefit from lower latency preemption
and re-scheduling.

Workloads like AMQP do asynchronous messaging, and are likely
to benefit from having a lower number of switches.

I do not know which kind of workload is more prevalent.

Another worry with gang scheduling is scalability.  One of
the reasons Linux scales well to larger systems is that a
lot of things are done CPU local, without communicating
things with other CPUs.  Making the scheduling algorithm
system-global has the potential to add in a lot of overhead.

Likewise, removing the ability to migrate workloads to idle
CPUs is likely to hurt a lot of real world workloads.

Benchmarks don't care, because they run full-out. However,
users do not run benchmarks nearly as much as they run
actual workloads...

--
All rights reversed
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/