[CentOS-devel] CentOS Cloud SIG?

Tue Jan 14 08:58:34 UTC 2014
Thomas Graf <tgraf at redhat.com>

On 01/13/2014 07:41 PM, Josh Boyer wrote:
>> We tried, quite a bit, to try and get the xen stuff backported into the
>> 2.6.32-EL6 kernel, but given the patch overlap was something we cant
>> control and lack the time and technical depth in kernel code to maintain
>> that longer term ( this was essentially Johnny and me, doing this over
>> and above most other things ), it was just easier to go down that route.

I'd be very interested to see a list of commits and corresponding
benchmarks for consideration into RHEL.

> OK.  Those kinds of reasons are what I would expect to drive a
> divergence from the disto kernel.  I'm curious to see how those kinds
> of situations work out in terms of bug reports and additional
> maintenance burden.
>
>> Secondly, there are quite a few anecdotal pieces on dramatic performance
>> improvements in newer kernels, I know atleast one of the top 5 CDN's (
>> ie carry > 25Gibt/sec ) that recently switched over from the distro
>> centos kernel to a inhouse 3.10 build for 'network perforamnce'.
>>
>> In quotes, because repeated questions and pokes in various media have
>> failed to get a reasonable, tangiable, technical response out of them
>> beyond 'network perforamance'.
>
> Interesting.  I suppose it's possible but without actually data
> there's really no way to tell.  It might even be that they modified
> 3.10 themselves to gain that performance and are reluctant to talk
> about it?

I suppose you don't have any additional details.

We did a ton of performance work for RHEL6.5 and are beating upstream
performance in various benchmarks while we have a minimal gap in a few
benchmarks. Some of them track down to a difference in how memory
accounting is performed. That is of course assuming a stock kernel
without any bypass technology such as DPDK or SR-IOV.

Just bumping to 3.10 or 3.12 is not going to be an universal performance
booster fix.