Date: Thu, 31 May 2007 11:09:18 +0900 From: "George V. Neville-Neil" <gnn@neville-neil.com> To: security <security@jim-liesl.org> Cc: freebsd-performance@freebsd.org, Andrew Gallatin <gallatin@cs.duke.edu> Subject: Re: [Fwd: asymetric speeds over gigE link] Message-ID: <m2ps4hoib5.wl%gnn@neville-neil.com> In-Reply-To: <465DE367.8070009@jim-liesl.org> References: <4652383E.9000302@jim-liesl.org> <18013.49538.356407.283631@grasshopper.cs.duke.edu> <465DE367.8070009@jim-liesl.org>
next in thread | previous in thread | raw e-mail | index | archive | help
At Wed, 30 May 2007 13:49:43 -0700, security wrote: > Drew, > Odd you should mention netperf. It yielded nearly identical > asymmetric results as iperf. I ended up testing with netpipe (NPtcp). > It reported fairly symmetric speeds between the two boxes when the send > and recv buffers were set the same, so that mystery is solved (maybe). > That being said I still wonder if I've too abstracted the net perfromance. > > The down side is that the best speed I got was around 362Mbps on > a gigE link (with netpipe). I had hoped for better. NetPIPE is my test of choice for such things, but I have not fully evaluated where it's performance bottlenecks may lie. It is most useful for causing "problems" to be found with networking code because unlike most tests it attempts to use odd sized packets. An interesting experiment would be to run gprof on NetPIPE to make sure that it was not the source of the inefficiency. If you try that can you post results? I'm the maintainer of the NetPIPE port in FreeBSD and have put in the code to cover IPv6 and SCTP. Best, George
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?m2ps4hoib5.wl%gnn>