I am running the supplied command-line throughput tools (local_thr and
remote_thr) on two machines running Windows XP SP3, connected via 100 Mb/s
Ethernet. I'm testing various message sizes, doing 3 test runs at each message
size, to see how tcp:// performance changes with message size. Everything goes
as I expect, up until I try a message size of 10000. Confused, I start doing a
binary search to find when things start being strange.
Message size of 8978: average throughput of 64.714 Mb/s
Message size of 8979: average throughput of 20.903 Mb/s
Message size of 8980: average throughput of 15.590 Mb/s
Message size of 8981: average throughput of 11.304 Mb/s
Message size of 8982: average throughput of 8.743 Mb/s
Message size of 8983: average throughput of 8.313 Mb/s
The change slows, until at a message size of 8990 it's mostly stabilized at
about 4 Mb/s.
Does anyone have any idea why this would happen? Neither CPU usage nor memory
usage are unusual during these runs. My only theory is that I'm hitting some
boundary in the memory cache, but that doesn't seem reasonable to me either.
Random system details which probably aren't relevant:
3 GHz Pentium 4 CPUs
512 MB memory each
Broadcom NetXtreme 57xx gigabit Ethernet cards
Connected via a Dell PowerConnect 2024 100 megabit switch
[email protected]<mailto:[email protected]>
(who doesn't understand why the router is only 100 megabit)
_______________________________________________
zeromq-dev mailing list
[email protected]
http://lists.zeromq.org/mailman/listinfo/zeromq-dev
