Date: Wed, 9 Dec 2015 08:13:44 -0800 From: Randall Stewart <rrs@netflix.com> To: FreeBSD Transports <transport@FreeBSD.org> Cc: hiren panchasara <hiren@strugglingcoder.info>, Jonathan Looney <jlooney@juniper.net>, "Robert N. M. Watson" <rwatson@freebsd.org> Subject: Preliminary results from testing.. Message-ID: <311DFF1D-A716-48B9-AB9C-B4A07DFF8936@netflix.com>
next in thread | raw e-mail | index | archive | help
All: Ok so I have had a first cut of an A/B test with a machine using fixed = calls to tcp_do_segment/output etc and one using the new pluggable pointer based stack. As expected I can=92t = really tell any real difference between the two machines. They are both Intel Ivy Bridge E5-2650L Intel CPU=92s. Running at = 1.7Ghz, 64G of memory and 10 physical cores. Here are png=92s of the per second cpu use and the overall serving = bandwidth (if anyone wants the raw numbers let me know). The two machines are basically the same, serving the same content. I have a few other pairs I will be trying this on.. though I don=92t = expect any real difference. Robert, how are you coming on investigating the overhead of function = pointers on non-intel devices ? I would like to commit the tcp-stack changes soon if possible :-) R -------- Randall Stewart rrs@netflix.com 803-317-4952
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?311DFF1D-A716-48B9-AB9C-B4A07DFF8936>