[openib-general] Re: [PATCH] ipoib tune rx ring size
Grant Grundler
iod00d at hp.com
Tue Jun 21 00:45:38 PDT 2005
On Mon, Jun 20, 2005 at 11:23:30PM +0300, Michael S. Tsirkin wrote:
> Grant, are you willing to test the patch and post what effect if any this has
> on netperf?
Conclusion: RX_RING @ 64 degrades netperf by about 1%.
I'm certain my runs are CPU bound.
I'm not certain that's true for your config.
Can you post netperf output?
Original SVN r2577 code running netperf 2.4.0:
(Sorry stats are busted on this verion of netperf - I'm tracking down that bug)
/usr/local/bin/netperf -l 60 -H 10.0.0.30 -t TCP_STREAM -T 1 -c 1024 -C 1024 -- -m 131072 -s 131072 -S 131072
Recv Send Send Utilization Service Demand
Socket Socket Message Elapsed Send Recv Send Recv
Size Size Size Time Throughput local remote local remote
bytes bytes bytes secs. 10^6bits/s % S % S us/KB us/KB
249856 249856 131072 60.00 2728.02 34.45 42.64 2.069 2.561
With IPOIB_RX_RING_SIZE at 64:
249856 249856 131072 60.00 2702.37 34.39 42.46 2.085 2.574
249856 249856 131072 60.00 2700.92 34.35 42.36 2.084 2.569
249856 249856 131072 60.00 2701.02 34.61 42.36 2.100 2.569
I'm satisfied the runs are fairly consistent.
More caveats:
Original run had CPU pinned to same CPU that was taking interrupts
for both ends of the connection.
I just happened to get the same CPU on "netperf" client side
and was pinning the netserver ("-T 1") side.
(If I didn't, I'd see ~3300-3500 Mb/s.)
However, binding the tasks to the non-interrupt CPU ("-T 0,0") results
in essentially no change in perf.
With RX_RING @ 128:
249856 249856 131072 60.00 3585.08 61.31 70.67 2.802 3.230
249856 249856 131072 60.00 3577.81 60.28 70.74 2.760 3.239
249856 249856 131072 60.00 3577.27 61.16 70.38 2.801 3.223
With RX_RING @ 64:
249856 249856 131072 60.00 3590.58 60.94 70.48 2.781 3.216
249856 249856 131072 60.00 3569.83 61.25 70.47 2.811 3.234
249856 249856 131072 60.00 3564.18 60.32 70.48 2.773 3.240
hth,
grant
More information about the general
mailing list