Date: Wed, 15 Mar 2006 09:18:06 +0900 From: Pyun YongHyeon <pyunyh@gmail.com> To: OxY <oxy@field.hu> Cc: freebsd-net@freebsd.org Subject: Re: problem with Marwell gigabit performance Message-ID: <20060315001806.GA52826@cdnetworks.co.kr> In-Reply-To: <001101c64754$cbbd65d0$0201a8c0@oxy> References: <001101c64754$cbbd65d0$0201a8c0@oxy>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Mar 14, 2006 at 11:48:14AM +0100, OxY wrote: > hi! > > i have a Marwell (SMC) gigabit ethernet card (sk0) and > have serious problems with performance.. > the machine is a pc, amd 2000+ xp, 512mb ram. > > tested with iperf (bidirectional test, udp transfer, not stream) > and got 8-15% packet drop when the system was idle. > There is a modified sk(4) driver at http://people.freebsd.org/~yongari/sk/sk_test/if_sk.c http://people.freebsd.org/~yongari/sk/sk_test/if_skreg.h Would you please give it try? I can't sure but you would get better results as the new driver uses one lock for standard MTU size and supports Tx TCP checksum offload and Rx IP checksum offload. I could push 32bit PCI NIC(DGE-530T) to the PCI bus limit on sparc64. > then tuned the sysctl with these settings: > > kern.ipc.maxsockbuf=81920000 > kern.ipc.nmbclusters=5000000000 > kern.ipc.somaxconn=8192 > net.inet.tcp.inflight.enable=1 > net.inet.raw.recvspace=4096 > net.inet.tcp.recvspace=131072 > kern.ipc.shmmax=409600000 > > with these i got around 3-7% packet drop, > but it's very high compared to zero :) > > my question is where/what should i optimize to not > have packet drop at all, or at least reduce it as much as > possible. > > thanks for your help! > > Csaba Banhalmi, HU -- Regards, Pyun YongHyeon
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20060315001806.GA52826>