Date: Sat, 20 Oct 2001 03:16:19 +0200 From: Marko Zec <zec@tel.fer.hr> To: Mike Tancsa <mike@sentex.net> Cc: freebsd-net@freebsd.org Subject: Re: fxp driver - receive interrupt bundling Message-ID: <3BD0D063.C084B21D@tel.fer.hr> References: <SEN.1003532776.274194794@news.sentex.net> <o1i1ttcg3fbn3n8n4mlv64ta9kldnmskld@4ax.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Mike Tancsa wrote: > > Hi, > What values would you reccomend for a box with 4 fxp NICs in them ? > Does it follow the the INT_DELAY should be lowered as you add more NICs ? > It doesn't matter how many fxp cards you have installed - if your box acts as a server than most probably you can leave INT_DELAY at default value (Intel proposes 0x600, but I think 0x400 would be more appropriate). If you use your multi-fxp-card BSD box as a router, than the microcode will impose additional delay *twice* (once in each direction), so in that case the default value of 0x600 might be too high for achieving full 100 megE throughput, because of TCP windowing scheme having to wait for ACK frames, which will be held in fxp receive buffers too long. On the other hand, setting INT_DELAY too low minimizes the benefits of bundling interrupts, as fewer received frames get "bundled" on a single interrupt. To summarize: if you are doing any routing (or bridging as I do), find the best value for INT_DELAY for your specific environment experimentally, it should be definitely smaller than or equal to 0x400. If you don't do packet forwarding between fxp interfaces, use the defaults. Marko 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?3BD0D063.C084B21D>