Mark, You're correct, we observed the same thing. The CPU/Load is caused by kswapd. Errors exactly like this has been reported for 2.4.2 and 2.4.9 (that we can find), so perhaps one of the bugs re-surfaced in 2.4.2X.
We've tried this with and without Herbert's rmap15k patches. Slightly different behavior is observed, but the end result is the same. I've just got finished a working port of vserver under UML (very cool!). Herbert is going to look it over/redo it and then I don't see why it wouldn't be released. I'm going to try to split up the load under multiple UML's (ie. 4-8 vservers per UML instance) and then I can get full console (with sysrq!). On Mon, 2003-11-17 at 13:00, Mark Lawrence wrote: > Hi Matt, > > On Mon, 17 Nov 2003, Matt Ayres wrote: > > I have 2 servers who keep crashing, sometimes every few hours... I am > > still working on a console solution so I don't know the exact place > > where the problem is. > > I can perhaps provide a little bit of information with regards to this, > since I am hosted on one of those machines :) > > You will see from the attached graph, that a symptom (although probably > not the cause) of each hang is a continual descreasing of free and > unbuffered memory. Swap seems to be untouched. > > I have also seen an increase of the cpu (up to 60-70%) used for "system" > tasks right before each hang. I bet if you extrapolate the current usage > you could reasonably predict the next hang... which at least make the > problem reproducable :) > > So it seems as if something in the virtual memory system is killing the > box, although how the vserver patch affect this I don't know because I > didn't think we actually touched the vm code. > > I would theorize that tcp connects are established ok, but that the first > thing to happen in most cases is to grab some free memory which doesn't > exist. Perhaps there is then some thrashing before things come to a > complete halt. > > Does anyone know what the normal kernel behaviour is with regards to > releasing cached/buffered memory in order to keep some free? What else > could cause this type of behaviour? > > Cheers, Mark. > -- > Mark Lawrence ([EMAIL PROTECTED]) -- Matt Ayres <[EMAIL PROTECTED]> TekTonic _______________________________________________ Vserver mailing list [EMAIL PROTECTED] http://list.linux-vserver.org/mailman/listinfo/vserver
