On 28.3.2018. 11:42, Hrvoje Popovski wrote:
> On 28.3.2018. 3:28, David Gwynne wrote:
>> On Thu, Mar 15, 2018 at 03:25:46PM +0100, Martin Pieuchot wrote:
>>> On 14/03/18(Wed) 13:00, David Gwynne wrote:
>>>> this adds transmit mitigation back to the tree.
>>>>
>>>> it is basically the same diff as last time. the big difference this
>>>> time is that all the tunnel drivers all defer ip_output calls, which
>>>> avoids having to play games with NET_LOCK in the ifq transmit paths.
>>> Comments inline.
>>>
>>>> + if (ifq_len(ifq) >= min(4, ifq->ifq_maxlen)) {
>>> Why 4? DragonFly recently bumped `ifsq_stage_cntmax' to 16. Did you
>>> try other values? They also have an XXX comment that this value should
>>> be per-interface. Why?
>> their default was 4, and they'd done some research on it. if they
>> moved to 16 there would be a reason for it.
> Hi all,
>
> with this diff i'm getting 1.43Mpps on
> 12 x Intel(R) Xeon(R) CPU E5-2620 v2 @ 2.10GHz, 2400.01 MHz
>
> with plain kernel i'm getting 1.12Mpps and with older diff's i was
> getting 1.31Mpps ...
>
> if i execute ifconfig ix0 down while generating traffic over everything stop
I've tested this diff with today's tree and i can't repeat problem with
or without -fretpoline diff.
