Re: Performance regression on kernels 3.10 and newer

[Rick Jones <rick.jones2@hp.com>]

On 08/14/2014 04:48 PM, Eric Dumazet wrote:
> Here is the thing
>
> nh_pcpu_rth_output is a per cpu cache.
>
> But nh_rth_input is not, because we did not optimize the case where dst
> has to be refcounted, yet.
>
> Rationale is explained in d26b3a7c4b3b26319f18bb645de93eba8f4bdcd5
> ("ipv4: percpu nh_rth_output cache")
>
> If you guys really believe we should have a percpu dst, go for it, but
> again, if the softirq handler runs on a different cpu than the
> application thread, it wont work.
>
> And, given 72 core servers are now on the way, we'll consume more ram,
> for netperf users.

Heck, we've had >= 72 core servers for *years* - go back to the 
Itanium-based Superdome servers for example, and I suspect some of their 
Power-based contemporaries.  They just weren't plentiful.  After that, 
there were/are the 8-socket x86 boxes from various vendors, and now the 
BL920s.

I will not claim that netperf represents all apps.  Neither will I claim 
it represents most apps.  But I won't accept that it represents no apps :)

And if there is a reasonably clean way to introduce epoll/poll/select 
into netperf I'm willing to consider it.

happy benchmarking,

rick jones