lkml.org 
[lkml]   [2010]   [Oct]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Subject[RFC PATCH 0/1] vhost: Reduce TX used buffer signal for performance
From
Date
This patch will change vhost TX used buffer guest signaling from one by
one to 3/4 ring size. I have tried different size, like 4, 16, 1/4 size,
1/2 size, and found that the large size is best for message size between
256 - 4K with netperf TCP_STREAM test, so 3/4 of the ring size is picked
up for signaling.

Tested both UDP and TCP performance with guest 2vcpu. The 60 secs
netperf run shows that guest to host performance for TCP.

TCP_STREAM

Message size Guest CPU(%) BW (Mb/s)
before:after before:after

256 57.84:58.42 1678.47:1908.75
512 68.68:60.21 1844.18:3387.33
1024 68.01:58.70 1945.14:3384.72
2048 65.36:54.25 2342.45:3799.31
4096 63.25:54.62 3307.11:4451.78
8192 59.57:57.89 6038.64:6694.04

UDP_STREAM
1024 49.64:26.69 1161.0:1687.6
2048 49.88:29.25 2326.8:2850.9
4096 49.59:29.15 3871.1:4880.3
8192 46.09:32.66 6822.9:7825.1
16K 42.90:34.96 11347.1:11767.4

For large message size, 60 secs run remains almost the same. I guess the
signal might not play a big role in large message transmission.

Shirley



\
 
 \ /
  Last update: 2010-10-27 23:03    [from the cache]
©2003-2011 Jasper Spaans