Re: [PATCH] bnx2x: add support for receive hashing

[Brian Bloniarz <bmb@athenacr.com>]

Eric Dumazet wrote:
> Le mardi 27 avril 2010 à 09:37 -0400, Brian Bloniarz a écrit :
>> David Miller wrote:
>>> How damn hard is it to add two 16-bit ports to the hash regardless of
>>> protocol?
>>>   
>> Come to think of it, for UDP the hash must ignore
>> the srcport and srcaddr, because a single bound
>> socket is going to wildcard both those fields.
>>
> 
> For your application maybe ;)
> 
> Here, I have thousand of RTP flows to big mediagateways, so the
> (srcaddr, dstaddr) is shared by all these flows.
> 
> Tom Herbert also wants a threaded DNS server.
> 
> For UDP, we could have a bitmap (system level ?) to say if a particular
> destination port wants multi-cpu (RPS) spreading or not, even if the NIC
> decided to use a single queue to submit frames to the host (ie mask the
> src_addr and src_port). In this case, RFS on non conected UDP sockets
> could be activated as well.
> 
> Or just a global sysctl to be able to mask the src_addr and/or src_port
> in our software rxhash.

This is all good to know. Here are 3 other alternatives/suggestions:

1) Leave things as they are, if it really hurts that badly
for our workload we can just disable RPS entirely.
2) Add a global sysctl to disable RPS for datagram-based protocols.
3) Pluggable SW rxhashes :)

I haven't benchmarked anything for our workloads yet, so
I can't say for sure it's even a big issue.

Has anyone benchmarked RPS + single-threaded DNS servers? It'd
be a pessimization in that case, right? Even the multi-threaded
DNS server wouldn't be workable unless there was something like
the soreuseport patch you & Tom had been been discussing.