The main goal of this series of patches is to reverse a regression in
netperf TCP_CRR performance results that was noticed in the transition
from Open vSwitch 1.2 to 1.3. The Netperf documentation describes
TCP_CRR as:
A TCP_CRR test is like a merger of a TCP_RR and TCP_CC test
which measures the performance of establishing a connection,
exchanging a single request/response transaction, and
tearing-down that connection. This is very much like what
happens in an HTTP 1.0 or HTTP 1.1 connection when HTTP
Keepalives are not used. In fact, the TCP_CRR test was added
to netperf to simulate just that.
Performance optimizations in the Open vSwitch 1.3 series focused on
cases in which many flows must be set up at a time (as measured by
ovs-benchmark). However, the TCP_CRR test keeps only a single
connection in flight at any given time. At any time, progress
requires getting a response to the packet sent most recently. The
performance for TCP_CRR actually dropped:
TCP_CRR ovs-benchmark
-------------- -------------
branch-1.2 1237/1047/1042 9063
branch-1.3 765/ 785/ 811 13310
(I report the values for three consecutive runs of TCP_CRR, since that
tends to have a lot of variance, but only one for a 20-second run of
ovs-benchmark because that is much more stable.)
This series improves matters. I originally developed it against
branch-1.3. The following are the numbers I measured for each commit
when it was on branch-1.3. (The baseline numbers differ from
branch-1.3 above because I ran them in slightly different environments.)
TCP_CRR ovs-benchmark
-------------- -------------
branch-1.3 baseline 885/ 840/ 854 12126
+ Remove poll_fd_woke()... 706/ 584/ 592 10832
+ Use poll() internally in... 884/ 851/ 818 12734
+ Use "epoll" instead of... 1083/ 999/ 998 12807
+ Process multiple batches... 1300/1295/1302 12440
+ Add "fast path". 1593/1580/1647 13716
------------------------------ -------------- -------------
improvement versus baseline +87% +13%
I ran the same tests on current "master" versus with all the patches
applied, without breaking it down by individual patch, and saw similar
improvements:
TCP_CRR ovs-benchmark
-------------- -------------
master commit 0145d7ed02095 868/ 813/ 800 12153
+ this series of patches 1527/1556/1553 13675
------------------------------ -------------- -------------
improvement versus master 0145 +87% +13%
improvement versus branch-1.2 +39% +51%
Ben Pfaff (6):
dpif-linux: Remove poll_fd_woke() optimization from
dpif_linux_recv().
Revert "poll-loop: Enable checking whether a FD caused a wakeup."
dpif-linux: Use poll() internally in dpif_linux_recv().
dpif-linux: Use "epoll" instead of poll().
ofproto-dpif: Process multiple batches of upcalls in a single poll
loop.
ofproto: Add "fast path".
lib/dpif-linux.c | 109 +++++++++++++++++++++++++++++--------------
lib/netlink-socket.c | 15 ++++--
lib/netlink-socket.h | 2 +-
lib/poll-loop.c | 33 ++-----------
lib/poll-loop.h | 3 -
ofproto/ofproto-dpif.c | 113 +++++++++++++++++++++++++++----------------
ofproto/ofproto-provider.h | 1 +
ofproto/ofproto.c | 6 ++
ofproto/ofproto.h | 1 +
vswitchd/bridge.c | 10 ++++
vswitchd/bridge.h | 3 +-
vswitchd/ovs-vswitchd.c | 2 +
12 files changed, 182 insertions(+), 116 deletions(-)
--
1.7.4.4
_______________________________________________
dev mailing list
[email protected]
http://openvswitch.org/mailman/listinfo/dev
