FreeBSD problems and preliminary ways to solve

[Robert N. M. Watson]

On 19 Aug 2011, at 10:23, Slawa Olhovchenkov wrote:

> On Fri, Aug 19, 2011 at 09:41:41AM +0100, Robert Watson wrote:
> 
>> There are a few known issues in terms of parallelism -- one is contention 
>> between the ithread and user thread on per-TCP connection locks.  Another is 
>> that we still haven't managed to switch to per-CPU statistics for the network 
>> stack (which is fairly straight forward in a specific sense, but we'd like a 
>> proper abstraction for it so we can generalise).
> 
> After discussion in other place: may be perfomace improve if
> application resheduling on the same core, as processing incoming
> IP packet (hot CPU caсhe)?

We have a GSoC student working on this currently (measuring rough socket affinity for applications a la RPS and redirecting work on the software side). I also have some related work, which can supplement his, which will push those application affinities into programmable TCAMs/hardware hash tables in 10gbs cards. We do need some general improvements in scheduling to allow the scheduler to take into account notions of vertically aligning affinity, however.

Robert