Date: Wed, 23 Jan 2002 12:44:37 -0500 From: Brian McGovern <mcgovern@spoon.beta.com> To: Lars Erik Gullerud <lerik@kq.no> Cc: qa@freebsd.org Subject: Re: 4.5-RC2 kernel, m_clalloc failed Message-ID: <200201231744.g0NHib912119@spoon.beta.com> In-Reply-To: Your message of "23 Jan 2002 15:58:52 %2B0100." <1011797933.75400.19.camel@elmer.i.eunet.no>
next in thread | previous in thread | raw e-mail | index | archive | help
In my opinion, I don't think you can really say something is broken, merely mis-tuned. From what I know, NMBCLUSTERS, when not explicitly set, is scaled based on MAXUSERS. So, what you've done is to say: MAXUSERS = some guessed-at number NMBCLUSTERS = MAXUSERS * some guessed-at order of magnitude for an "average" "user" (note the individual quotes, meaning that average and user need to be defined differently) NMBCLUSTERS is less than the number you need for your application. So, perhaps you should look at the number of NMBCLUSTERS on a working machine (information that in my read through appers to be missing), and compare it to the auto-calculated values. If they're really close, it might be a real problem (or just a marginal case). If they're significantly different, than what you're saying is that auto tuning doesn't work in YOUR case (I happy run with MAXUSERS between 64 and 256 on all machines, and have _never_ had a problem). I expect the reason for the self-tune was for newer boxes, 2+GHz monsters with 2-4GB of RAM and a Gig-E port that were being bottlenecked by the fact that MAXUSERS (and all of its derived values) were being set to a number lower than the sweet spot by being fixed, and the auto-tune code makes a stab at doing something better in the _average_ case. > I moved one of our shoutcast streaming servers over to 4.5-RC2 during > the weekend to test it in a high network load environment. It was > installed "fresh" on a clean system with an FTP install, i.e. not an > upgrade from the old 4.4 install. > > After being in operation for just a few hours the server started logging > these messages continously: > > Jan 23 14:34:41 disrv01 /kernel: m_clalloc failed, consider increase > NMBCLUSTERS value > Jan 23 14:34:41 disrv01 /kernel: fxp0: cluster allocation failed, packet > dropped! > Jan 23 14:34:42 disrv01 last message repeated 850 times > > The server had to be rebooted in order to function properly again, > network connectivity was basically gone until it was rebooted. After > reboot it again functioned for a few hours and then the same thing > happened. > > The server is running 5 instances of the shoutcast server (original > sc_serv binary from shoutcast, not icecast) on 5 separate TCP ports. The > load had been steady between 400 and 500 concurrent TCP streams @ > 128kbit for the whole 24 hours, and was at apx. 450 streams served when > this error hit. Nothing else is running on this server. > > Kernel was recompiled for this machine after install, so not a GENERIC > kernel, maxusers was set to 0 to test the auto-allocation. The > NMBCLUSTERS as reported in kern.ipc.nmbclusters=8640. This is a P3-800 > 512MB RAM box, who has been running steadily at a lot higher loads than > this for a long time on 4.4-STABLE (750 streams max), so I'm assuming > it's an issue introduced in 4.5. Unless it's just my stupidity and > something just needs to be configured differently - maybe I should avoid > the maxusers=0 and set some table sizes manually or something? > > I don't know what kind of info you might require, I'm not really much of > a "kernelhacker", but I hope you fix whatever the prob is before > 4.5-RELEASE or I'll have to stay on 4.4 for our streaming boxes. > > Regards, > Lars Erik Gullerud > > > > To Unsubscribe: send mail to majordomo@FreeBSD.org > with "unsubscribe freebsd-qa" in the body of the message To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-qa" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200201231744.g0NHib912119>