[netperf-dev] Controlling socket priority

[Rick Jones]

On 08/24/2011 11:43 AM, Dave Taht wrote:
> I am curious as to whether this really mucks with the TOS bits at all,
> especially on ipv6. I think it just mucks with the internal queuing.

have you tried Amir's patch on your copy of netperf?  I've not applied 
it to the top of trunk yet (BTW, Amir, I may not apply it to the 
"classic" tests just the omni path - I'm trying to migrate things to the 
"omni" code.) but I have pushed a doubling of the control message to 512 
bytes, guaranteeing that TOT will not work with anything previous. 
After that percolates for a couple days I will apply Amir's patch.

> I tried various things to get ipv6 to do the right thing, perhaps I
> was on drugs? See:
>
> http://www.bufferbloat.net/issues/249

Sounds like cause to check the netdev archives and then perhaps send an 
email to netdev.

happy benchmarking,

rick jones

BTW, the last time I altered the size of the control message was between 
versions 2.0 and 2.1.  Perhaps I should actually bump it to 1024 bytes 
rather than 512 and perhaps have it take even longer to need to do it 
again :)

>
> On Wed, Aug 24, 2011 at 11:15 AM, Rick Jones<rick.jones2 at hp.com>  wrote:
>> On 08/14/2011 05:29 AM, Amir Vadai wrote:
>>>
>>> Attached a fixed patch.
>>>
>>> Now it will also compile when SO_PRIORITY is not defined.
>>
>> Thanks. You have been very patient and I appreciate it.
>>
>> Now we come to the fun part.  By my calculations, the changes I made in the
>> top of trunk to actually report back transport layer retransmissions from
>> netserver's side occupied the last four bytes of the omni_response_struct or
>> more accurately, the overall 256 byte netperf control message size limit.
>>
>> So, that means bumping the size of the control message.  I've never held
>> that there would be inter-version support in netperf anyway but it has been
>> a long time since the last increase in control message size (at least I
>> think I've had a control message size increase in the past).
>>
>> rick jones
>>
>
>
>