Re: [Xen-devel] [for-4.8 PATCH] xen: credit2: make runqueues be per-socket by default.

[Wei Liu <wei.liu2@xxxxxxxxxx>]

On Tue, Nov 29, 2016 at 11:57:03AM +0100, Dario Faggioli wrote:
> Benchmarks have shown that per-socket runqueues arrangement
> behaves better (e.g., we achieve better load balancing)
> than the current per-core default.
> 
> Here's an example (coming from
> https://lists.xen.org/archives/html/xen-devel/2016-06/msg02287.html ):
> 
> |=======================================|
> | XEN BUILD TIME, LOW LOAD, NO NOISE    |
> |---------------------------------------|
> |       runq=core   runq=socket         |
> |         35.200       33.433           |
> |---------------------------------------|------------------------------|
> | XEN BUILD TIME, HIGH LOAD, NO NOISE   | IPERF, HIGH LOAD, NO NOISE   |
> |---------------------------------------|------------------------------|
> |       runq=core   runq=socket         |     runq=core runq=socket    |
> |         18.013       18.530           |       23.200     23.466      |
> |---------------------------------------|------------------------------|
> | XEN BUILD TIME, LOW LOAD, WITH NOISE  |
> |-------------------------------------  |
> |       runq=core   runq=socket         |
> |         45.866       39.493           |
> |---------------------------------------|------------------------------|
> | XEN BUILD TIME, HIGH LOAD, WITH NOISE | IPERF, HIGH LOAD, WITH NOISE |
> |---------------------------------------|------------------------------|
> |       runq=core   runq=socket         |     runq=core runq=socket    |
> |         36.840       29.080           |       19.967     21.000      |
> |=======================================|==============================|
> 
> The only reason why we went for per-core, initially, was to
> introduce some form of hyperthreading support. Now we have
> hyperthreading support, independently from how runqueues
> are organized (9bb9c7388 "xen: credit2: implement true SMT
> support"), and thus we can switch to per-socket.
> 
> Signed-off-by: Dario Faggioli <dario.faggioli@xxxxxxxxxx>
> Acked-by: George Dunlap <george.dunlap@xxxxxxxxxxxxx>
> ---
> Cc: Anshul Makkar <anshul.makkar@xxxxxxxxxx>
> Cc: Wei Liu <wei.liu2@xxxxxxxxxx>
> ---
> George's Ack obtained 'out of band', as he is away.
> ---

Strictly speaking, this is not necessary for 4.8 because even though the
default configuration is not optimal, there is a simple way to change
it.

On the other hand, because credit2 is declared supported in this
release, it is better to have a sensible default setting for credit2.

The risk is that changing the default value would cause osstest to
discover bugs in the scheduler logic, hence blocking the release. But
overall I think the chance is low, but Dario can you please state to
what extend did you test this patch?

Wei.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel