Date: Wed, 20 Jan 1999 16:22:08 -0500 From: Brian Del Vecchio <bdv@parlez.com> To: Chrisy Luke <chrisy@flix.net>, freebsd-stable@FreeBSD.ORG Cc: wolfgang@wsrcc.com Subject: Re: Cisco/Intel Ethernet Trunking Message-ID: <36A64900.7D2CF5C@parlez.com> References: <Pine.BSF.4.01.9901151325050.492-100000@newton.aipo.gov.au> <carl@xena.IPAustralia.gov.au> <9901142208.ZM5749@beatrice.rutgers.edu> <19990120195237.B25008@flix.net>
next in thread | previous in thread | raw e-mail | index | archive | help
Chrisy Luke wrote: > Allen Smith wrote (on Jan 15): > > See ftp://ftp.flirble.org/pub/unix/hacks/FreeBSD/mpath. This work is > > still in progress; one thing that would be good for this would be > > doing packet output to the least-loaded interface, instead of via the > > current round-robin method. > > When I get time that is certainly one thing that will be an optional > feature. MY idea is to measure the current buffered queue length (there > is such a thing) - shift by some metric specifyable per next-hop and > round-robins the group of the same and smallest number. If the metric is > 0, it presets the result of the calc to zero. If all metrics are zero > it by inference turns off the load-balancing. An alternative to consider is to randomly distribute packets across the equal-cost links. This may give you distribution that's as good as or better than any fancy queue-depth metrics. As may have already been discussed here, or may be well-known, the main risk of load sharing (according to my friend Wolfgang) is that you may end up getting TCP segments received out of order. For some implementations of TCP, this will result in segments being discarded and retransmitted, a noticable and detrimental side effect. brian To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-stable" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?36A64900.7D2CF5C>