On Fri, May 20, 2005 at 04:59:44PM +0200, Bernhard Fischer wrote: > Hi, > > On i386 UP, i see odd results during runs. The machine is basically > idle.
The runs are really short if you are using the default (-n 1000). I don't expect them to be visible to something like top or vmstat. > The worst latency is always somewhere around exchange 12-80. > Is this expected and if not, does somebody have an idea what may cause > this? It probably depends on which x86 chipset you are using and what else might be going on. > Thank you, > > $ ./rdma_lat -i 2 -U -s 1 tst09 -t 1 | egrep -v ", 3.(1|2|3)" > local address: LID 0x02, QPN 0x1a0406, PSN 0x5379e1 RKey 0x1040436 > VAddr 0x00000008051001 > remote address: LID 0x01, QPN 0x1d0406, PSN 0xd7aed6, RKey 0x188043c > VAddr 0x00000008051001 > #, usec > 1, 12.8368 Interesting that this is only ~4x normal latency. > 2, 4.94677 > 12, 7.0644 > 13, 87.8933 > 26, 64.8241 Could it be we are crossing a cacheline boundary someplace every 13 or 14 exchanges? Any simple statistical method to determine if the long delays are cyclical? grant _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
