polling's future [was: Re: Dynamic Ticks/HZ]

From: Andre Oppermann <andre_at_freebsd.org>
Date: Tue, 06 Nov 2012 11:23:34 +0100
On 05.11.2012 17:57, Luigi Rizzo wrote:
> On Mon, Nov 05, 2012 at 04:25:36PM +0000, Joe Holden wrote:
>> Luigi Rizzo wrote:
>>> On Mon, Nov 05, 2012 at 08:11:41AM -0500, Ryan Stone wrote:
>>>> On Mon, Nov 5, 2012 at 4:40 AM, Joe Holden <lists_at_rewt.org.uk> wrote:
>>>>
>>>>> doh, running kernel wasn't as GENERIC as I thought it was, looks like
>>>>> device polling not only breaks dynamic ticks but also reduces rx ability
>>>>> significantly, exactly 150,000 pps per 1000hz on igb versus 650,000
>>>>> without
>>>>>
>>>>> Is this a known issue? (and if device polling isn't as useful as it once
>>>>> was, should it be removed?)
>>>>>
>>>> Device polling on modern multiqueue NICs isn't very useful because you're
>>>> limited to a single thread for handling packets.  I have a patch that
>>>> fixes
>>>> this that I've let fall by the wayside.
>>>
>>> the 150,000 is result of the combination of the default value of
>>> sysctl  kern.polling.burst_max and kern.polling.idle_poll=0
>>> (i think this is the default value for the latter).
>>>
>>> The 150 was sized for the peak pps on a 100Mbit/s interface,
>>> back in 2001. You should at least be able to raise the number
>>> and see what kind of throughput you can achieve.
>>>
>>> This said, modern nics also have interrupt moderation so you
>>> don't really need polling.
>>>
>>> cheers
>>> luigi
>> Hi Luigi,
>>
>> This makes sense, am I likely to achieve better throughput (in the
>> forwarding path at this point) with netisr rather than polling,
>> especially as mentioned above the igb does indeed have multiple queues
>> for rx?
>
> at 1Gbit/s you probably don't need multiqueue (I am actually surpised
> you can only do 650kpps, but perhaps because you are using ipfw and
> not just doing plain forwarding ?)

Hi Luigi,

do you agree on polling having outlived its usefulness in the light
of interrupt moderating NIC's and SMP complications/disadvantages?

And also that in its current state it is providing no benefit, if
not negative, and that fixing it for the current world-order is major
undertaking that's not really beneficial considering superior
alternatives?

That nobody is maintaining it and taking care of and fixing the frequent
problem, performance and bug reports?

Additionally it misleads people without deep network knowledge to
compile it into their kernel under wrong assumptions and see a
degradation of performance, if not other side effects?

Under these considerations I propose to remove polling support from
current and 10.0 in lieu of upcoming superior alternatives (netmap
enabled features).  It was a great feature in the past but is beyond
retirement and should live a peaceful live in the attic.

-- 
Andre
Received on Tue Nov 06 2012 - 09:22:05 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:32 UTC