lkml.org 
[lkml]   [2013]   [Jun]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH RFC ticketlock] Auto-queued ticketlock
From
Date
On Tue, 2013-06-11 at 10:53 -0700, Davidlohr Bueso wrote:

> I hate to be the bearer of bad news but I got some pretty bad aim7
> performance numbers with this patch on an 8-socket (80 core) 256 Gb
> memory DL980 box against a vanilla 3.10-rc4 kernel:

This doesn't surprise me as the spin lock now contains a function call
on any contention. Not to mention the added i$ pressure on the embedded
spinlock code having to setup a function call.

Even if the queues are not used, it adds a slight overhead to all
spinlocks, due to the code size increase as well as a function call on
all contention, which will also have an impact on i$ and branch
prediction.


>
> * shared workload:
> 10-100 users is in the noise area.
> 100-2000 users: -13% throughput.
>
> * high_systime workload:
> 10-700 users is in the noise area.
> 700-2000 users: -55% throughput.
>
> * disk:
> 10-100 users -57% throughput.
> 100-1000 users: -25% throughput
> 1000-2000 users: +8% throughput (this patch only benefits when we have a

Perhaps this actually started using the queues?

> lot of concurrency).
>
> * custom:
> 10-100 users: -33% throughput.
> 100-2000 users: -46% throughput.
>
> * alltests:
> 10-1000 users is in the noise area.
> 1000-2000 users: -10% throughput.
>
> One notable exception is the short workload where we actually see
> positive numbers:
> 10-100 users: +40% throughput.
> 100-2000 users: +69% throughput.

Perhaps short work loads have a cold cache, and the impact on cache is
not as drastic?

It would be interesting to see what perf reports on these runs.

-- Steve




\
 
 \ /
  Last update: 2013-06-11 20:41    [W:0.199 / U:0.048 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site