[rfc] tcp timer update for RSS
adrian at freebsd.org
Sat May 17 14:52:30 UTC 2014
.. and as a note - it'll all be behind #ifdef RSS.
On 17 May 2014 07:49, Adrian Chadd <adrian at freebsd.org> wrote:
> On 17 May 2014 07:44, Bentkofsky, Michael <MBentkofsky at verisign.com> wrote:
>> Hi Adrian,
>> I haven’t had the chance to look this over carefully yet as we’re at BSDCan.
>> I think I understand what you’re trying to achieve by aligning the per-CPU
>> timer processing per core. In principal that sounds reasonable, although I
>> am unsure if you were trying to solve a particular performance issue with
>> this particular change. My sense is this is all preparatory with the goal of
>> all inp processing to become per core. Could you comment on the general
>> evolution you’re considering? Do most of the PCB structures become per-core,
>> as in PCB groups?
>> If you’d like us to test this change, I’m happy to do so. At the moment I
>> don’t know if we’d expect to see any benefit – do you have any traffic
>> conditions for which this showed any difference? But we can certainly drive
>> many hundreds of thousands of connections at reasonably high connection
>> rates if that will help.
> Yes, the aim is to provide RSS support in the RSS model of "align
> everything to a specific core." The goal for RSS is to remove both
> lock contention between cores and keep data CPU/cache local to improve
> efficiency there.
> There's nothing obvious that'll be beneficial right now. The items at
> https://wiki.freebsd.org/NetworkRSS have to occur before it is
> beneficial to anyone.
More information about the freebsd-net