SCTP performance?

SCTP performance?

[Ben Greear]

I just added some support for SCTP to my traffic-generator tool.

I tried to send 1Gbps of traffic, bi-directional, across two
10G interfaces.  TCP does this without any trouble at all,
but SCTP was asymetric and never got above around 700Mbps,
and was usually stuck at < 500Mbps.

perf top didn't show any obvious CPU bottlenecks.

Has anyone done any recent performance testing with SCTP?

I'd be interested to know the results if so...

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com

Re: SCTP performance?

[Eric Dumazet]

On Fri, 2013-08-09 at 16:11 -0700, Ben Greear wrote:
> I just added some support for SCTP to my traffic-generator tool.
> 
> I tried to send 1Gbps of traffic, bi-directional, across two
> 10G interfaces.  TCP does this without any trouble at all,
> but SCTP was asymetric and never got above around 700Mbps,
> and was usually stuck at < 500Mbps.
> 
> perf top didn't show any obvious CPU bottlenecks.
> 
> Has anyone done any recent performance testing with SCTP?
> 
> I'd be interested to know the results if so...

You mean there is no SCTP offload on the NIC and/or GRO, yes that's too
bad.

Re: SCTP performance?

[Ben Greear]

On 08/09/2013 04:21 PM, Eric Dumazet wrote:
> On Fri, 2013-08-09 at 16:11 -0700, Ben Greear wrote:
>> I just added some support for SCTP to my traffic-generator tool.
>>
>> I tried to send 1Gbps of traffic, bi-directional, across two
>> 10G interfaces.  TCP does this without any trouble at all,
>> but SCTP was asymetric and never got above around 700Mbps,
>> and was usually stuck at < 500Mbps.
>>
>> perf top didn't show any obvious CPU bottlenecks.
>>
>> Has anyone done any recent performance testing with SCTP?
>>
>> I'd be interested to know the results if so...
>
> You mean there is no SCTP offload on the NIC and/or GRO, yes that's too
> bad.
>

I'm sure that is part of the problem.

I got much better results when I explicitly set the
tx/rx socket buffer sizes to 2MB, so maybe it just sucks
at auto-tuning it's socket buffer sizes.

Thanks,
Ben

-- 
Ben Greear <greearb@candelatech.com>
Candela Technologies Inc  http://www.candelatech.com

Re: SCTP performance?

[David Miller]

From: Ben Greear <greearb@candelatech.com>
Date: Fri, 09 Aug 2013 17:04:59 -0700

> I got much better results when I explicitly set the
> tx/rx socket buffer sizes to 2MB, so maybe it just sucks
> at auto-tuning it's socket buffer sizes.

TCP is the only protocol in our stack which does socket buffer
size autotuning.

Re: SCTP performance?

[Vlad Yasevich]

On 08/09/2013 08:04 PM, Ben Greear wrote:
> On 08/09/2013 04:21 PM, Eric Dumazet wrote:
>> On Fri, 2013-08-09 at 16:11 -0700, Ben Greear wrote:
>>> I just added some support for SCTP to my traffic-generator tool.
>>>
>>> I tried to send 1Gbps of traffic, bi-directional, across two
>>> 10G interfaces.  TCP does this without any trouble at all,
>>> but SCTP was asymetric and never got above around 700Mbps,
>>> and was usually stuck at < 500Mbps.
>>>
>>> perf top didn't show any obvious CPU bottlenecks.
>>>
>>> Has anyone done any recent performance testing with SCTP?
>>>
>>> I'd be interested to know the results if so...
>>
>> You mean there is no SCTP offload on the NIC and/or GRO, yes that's too
>> bad.
>>
>
> I'm sure that is part of the problem.
>
> I got much better results when I explicitly set the
> tx/rx socket buffer sizes to 2MB, so maybe it just sucks
> at auto-tuning it's socket buffer sizes.
>
> Thanks,
> Ben
>

There is no buffer autotunning.  There is also no GRO since SCTP doesn't
lend itself easily to GRO. :(

There has been some research efforts at autotunning, but nothing 
proposed.  I'll see if I can find the code and make it work better.

-vlad