Date: Tue, 20 Aug 2002 16:43:28 -0400 (EDT) From: Andrew Gallatin <gallatin@cs.duke.edu> To: Luigi Rizzo <rizzo@icir.org> Cc: freebsd-net@FreeBSD.ORG Subject: Re: m_getcl and end-to-end performance Message-ID: <15714.43504.493596.791872@grasshopper.cs.duke.edu> In-Reply-To: <20020820132957.B49141@iguana.icir.org> References: <15714.27671.533860.408996@grasshopper.cs.duke.edu> <20020820093939.B48541@iguana.icir.org> <15714.39494.661931.882244@grasshopper.cs.duke.edu> <20020820132957.B49141@iguana.icir.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Luigi Rizzo writes: > > > For a netperf UDP_STREAM, I see ~1,000 to ~2,000 packets/sec increase > > in throughput for sizes 256 through 2K. For large (8K) I see a > > hmmm... that seems to be in the noise (i assume you are dealing > with a fast machine), but probably because there is a lot of > per-packet overhead in userland. In my experiments with routing of > 64-byte packets i saw the throughput bumping up from 360kpps to > over 400kpps on a fast box with Gig-E card, and about 2-3kpps (20 > to 22-23kpps) on the soekris box. This an SMP kernel using interrupts, not POLLING. And using the "-stable" version of our driver, which does *very* heavyweight interrupt processing and does not do any interrupt coalescing. There's a lot of overhead here, and much of it is ours. All of the is fixed in our "-current" driver.. Anyway, things max out at ~65K pkts/sec, so a few thousand pkts/second is a little better than in the noise. For a driver with lower overhead (fxp), I see ~12K pkts/sec improvemnt again, SMP and not POLLING) from ~90Kpkts/sec -> to 102Kpkts/sec for minimally tiny packets. Drew To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-net" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?15714.43504.493596.791872>