Skip site navigation (1)Skip section navigation (2)
Date:      Thu, 4 Aug 2011 19:08:37 +0900
From:      Kazuya Goda <gockzy@gmail.com>
To:        soc-status@freebsd.org
Subject:   [status report] RPS/RFS #week10
Message-ID:  <CA%2Bz95aVNQ-cgcWgdx1NC1fSH_9aNLL8qErMq2dgSpX8tXZwa-A@mail.gmail.com>

next in thread | raw e-mail | index | archive | help
Hi,

The project goal is to implement RPS/RFS on FreeBSD. RPS solves the
problem of mono-queue NIC which can't distribute packets across
multiple processors. RFS is the extension of RPS, which deliver
packets to the CPU where application running.


This week status:
* Research "lock" problem
The case of set below :
-net.isr.direct=0
-net.isr.direct_force=0
it cause to netisr(protocol stack) thread and dispatcher thread are
running on same CPU.
For the time begin net.isr.direct_force is set 1.


* Benchmark test
I used netperf. The netperf test has 50 instances of netperf TCP_RR
test with 1 byte request and response.
Below is result. The result SOFT_RSS is almost the same as RPS.
SOFT_RSS work on more  flows.
I'd like to benchmark with many flows but case of many flows drop
performance due to lock in protocol stack.

NO RPS/SOFT_RSS - 87k tps
RPS                         - 100k tps
SOFT_RSS               - 99k tps


next week :
* performance up

--Kazuy Goda



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?CA%2Bz95aVNQ-cgcWgdx1NC1fSH_9aNLL8qErMq2dgSpX8tXZwa-A>