Date: Fri, 10 Feb 95 9:02:53 MST From: terry@cs.weber.edu (Terry Lambert) To: wollman@halloran-eldar.lcs.mit.edu (Garrett Wollman) Cc: rminnich@mini.sarnoff.com, freebsd-questions@FreeBSD.org Subject: Re: enet throughput Message-ID: <9502101602.AA11205@cs.weber.edu> In-Reply-To: <9502100339.AA02532@halloran-eldar.lcs.mit.edu> from "Garrett Wollman" at Feb 9, 95 10:39:30 pm
next in thread | previous in thread | raw e-mail | index | archive | help
> Two almost identical Pentium 60's running yesterday's kernel, one with > a genuine SMC 8416 and one with a cheap 8216 clone: > > ttcp-r: 16777216 bytes in 16.56 real seconds = 989.23 KB/sec +++ > ttcp-r: 16777216 bytes in 2.89 CPU seconds = 5678.54 KB/cpu sec > ttcp-r: 11460 I/O calls, msec/call = 1.48, calls/sec = 691.93 > ttcp-r: 0.0user 2.8sys 0:16real 17% 52i+613d 138maxrss 0+2pf 6787+108csw > > ttcp-t: 16777216 bytes in 16.67 real seconds = 983.04 KB/sec +++ > ttcp-t: 16777216 bytes in 3.65 CPU seconds = 4494.08 KB/cpu sec > ttcp-t: 2048 I/O calls, msec/call = 8.33, calls/sec = 122.88 > ttcp-t: 0.0user 3.5sys 0:16real 21% 35i+440d 178maxrss 0+2pf 3969+41csw > > This on a moderately loaded (~ 100 machines) Ethernet. Hmmm... I think that ~5 times the number of I/O calls at ~1/6 the time per call is an interesting statistic. Seems to point right at the problem spots... minus one piece of discriminating information -- was this a UDP NFS problem or a TCP NFS problem? Latency hits hard (per packet) for request/response, and ttcp does not test for this, which yields one latency averaged across all packets. NFS writes which are not async are very nearly request/response because of the relative difference in time for window fill vs. doing a sync I/O to a disk. Terry Lambert terry@cs.weber.edu --- Any opinions in this posting are my own and not those of my present or previous employers.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?9502101602.AA11205>