Date: Fri, 13 Jul 2001 13:14:07 -0700 (PDT) From: Julian Elischer <julian@elischer.org> To: Leo Bicknell <bicknell@ufp.org> Cc: freebsd-hackers@freebsd.org Subject: Re: Network performance roadmap. Message-ID: <Pine.BSF.4.21.0107131302240.69775-100000@InterJet.elischer.org> In-Reply-To: <20010713101107.B9559@ussenterprise.ufp.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, 13 Jul 2001, Leo Bicknell wrote: > [...] I think you have analysed it very well.. I can summarise: We want the ability to make the current window to be able to grow to a very large value if the link needs it. So sndbuf shold eb set high. We don't want every process buffering in advance the entire potential maximum window size of data, but rather some safety margin over what is likely to be needed in a hurry by the tcpstack with it's instantaneous window. We don't want to make processes stop and start too much buffering small ammounts of data. I would suggest that 2 * the current window may be too small because the window might be increasing and reception of one ack might move the window up by the entire transmitted window size, resulting in starvation if the waiting data can be sent quickly. therefore I suggest a few modifying factors: A minimum value of something like 16K use 3* cWIN rather than 2* as a lower bound. also try figure out how much data we can send in 0.1 sec and use that as part of the calculation. Some systems with LOTS of server threads don't want to keep scheduling them but would rather have a minimum amount of data queued to allow fairly 'chunkuy' scheduling. To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.BSF.4.21.0107131302240.69775-100000>