On Wednesday 17 April 2013 19:41:38 José Orlando Pereira wrote: > That's what I have been doing. Using CS_DISPATCH_ONE in a loop and confirming that > there are no errors reported. Without the patch: > > 1274216 messages received 64 bytes per write 10.000 Seconds runtime 127420.504 TP/s 8.155 MB/s. > 913165 messages received 320 bytes per write 10.000 Seconds runtime 91313.158 TP/s 29.220 MB/s. > 388160 messages received 1600 bytes per write 10.000 Seconds runtime 38815.282 TP/s 62.104 MB/s. > 119352 messages received 8000 bytes per write 10.000 Seconds runtime 11935.158 TP/s 95.481 MB/s. > 31492 messages received 40000 bytes per write 10.000 Seconds runtime 3149.188 TP/s 125.968 MB/s. > 7520 messages received 200000 bytes per write 10.000 Seconds runtime 752.097 TP/s 150.419 MB/s. > 1446 messages received 1000000 bytes per write 10.000 Seconds runtime 144.599 TP/s 144.599 MB/s. > > With the patch: > > 534997 messages received 64 bytes per write 10.000 Seconds runtime 53498.500 TP/s 3.424 MB/s. > 532639 messages received 320 bytes per write 10.000 Seconds runtime 53263.052 TP/s 17.044 MB/s. > 516788 messages received 1600 bytes per write 10.000 Seconds runtime 51678.919 TP/s 82.686 MB/s. > 501374 messages received 8000 bytes per write 10.000 Seconds runtime 50137.505 TP/s 401.100 MB/s. > 517638 messages received 40000 bytes per write 10.000 Seconds runtime 51763.934 TP/s 2070.557 MB/s. > 152837 messages received 200000 bytes per write 10.000 Seconds runtime 15283.630 TP/s 3056.726 MB/s. > 1343 messages received 1000000 bytes per write 10.000 Seconds runtime 134.299 TP/s 134.299 MB/s. > > I'm still trying to understand why. I think I finally understood it. Unfortunately, no magic... :-) What seems to be happening is that 64 byte messages pile up, and are delivered only when the benchmark is already expecting 320 byte messages, and so on. This means that bandwidth as printed is wrong, as it assumes that messages being delivered have been sent with the corret size. This is already happening without my patch, but by introducing some overhead, it makes it (a lot) worse. Regards, -- Jose Orlando Pereira
Attachment:
smime.p7s
Description: S/MIME cryptographic signature
_______________________________________________ discuss mailing list discuss@xxxxxxxxxxxx http://lists.corosync.org/mailman/listinfo/discuss
