[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH, RFC] x86/HVM: batch vCPU wakeups

>>> On 09.09.14 at 23:29, <tim@xxxxxxx> wrote:
> At 09:33 +0100 on 09 Sep (1410251617), Jan Beulich wrote:
>> RFC for two reasons:
>> 1) I started to consider elimination of the event-check IPIs in a more
>>    general way when MONITOR/MWAIT is in use: As long as the remote CPU
>>    is known to be MWAITing (or in the process of resuming after MWAIT),
>>    the write of softirq_pending(cpu) ought to be sufficient to wake it.
>>    This would yield the patch here pointless on MONITOR/MWAIT capable
>>    hardware.
> That's a promising line.

And it alone gives quite a bit better performance than batching, at
least on the hardware I can directly test on. The maxima go down
even more significantly than the average values.

Putting the batching back on top improves the average delivery
times a little further, but surprisingly makes the maxima go back
up to almost where they were with just batching. So far I haven't
been able to think of an explanation for this. But while attaching
the raw patch here for reference, I'll want to wait with formal
submission until I've seen the numbers from the big box (also to
decide whether to include or drop the batching one).


Attachment: x86-suppress-event-check-IPI.patch
Description: Text document

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.