Re: memory management things to hack...

Koni (
Tue, 18 Nov 1997 15:39:42 -0500

Rik van Riel-

2.1.64 test sounds good, I will attempt to take control of the machine this
weekend if possible. I will send you any observed results. I have a
simple-minded program that I wrote which tries to hog up as much memory as
possible up to 260M and tests walking-bit patterns on the memory as it is
obtained. I usually "stress-test" with 3 or 4 of these progs running
concurrently at first. Upon successful execution of several of these tests,
then I let our model loose on it, which will hog memory, CPU, and disk
resources for hours. After that, the staff will come in the next day and
abuse the hell out of "misic" as per usual, so I guess needless to say, we
can provide substantial stress testing, even in the context of "normal" real
life workload.

Thanks to all who have responded with suggestions, ideas, and explainations.
Even if I never get it working to total satisfaction of our programmer and
analyst staff, I like at least understanding the situation!


A couple other notes, as some have inquired further details:

motherboard: Supermicro P6DNF, 2 Ppro-200 (256K cache) (400W power supply)
cards: a 3com905 and a 3com 509, Buslogic 958 UW-SCSI, Number Nine FX 771
mem chips, 8 64M EDO (generic) - all have been tested and verified to be
working correctly. Several combinations ( 8 choose 4 possible), ie 256M, have
been verified to work as well as any other combination of memory (even 8 32M
chips) that equals 256M. However, 512M has only been kosher with 2.1.62 so
far. (Of the 2.1.X series, I think i've tried a while ago 2.1.29 and some
others, which didn't work. Perhaps I should verify that if I get a chance)

Rik van Riel wrote:

> On Mon, 17 Nov 1997, Koni wrote:
> > To the person who requested mem management things to hack:
> > (I don't have your address I think because I joined the list too late)
> That's me...
> > I have an SMP machine (2 PPro cpus) with 512M (well, I'd like to say it
> > has 512M). I can confirm for anyone who wants to know that 2.0.30 works
> > fine with 256M of ram. However, it crashes almost immediately (during
> > boot) when the machine has 512M of ram installed. It doesn't even work
> > if I use "mem=256M" ( i don't know why). Now, I have managed to get the
> > machine usable with 512M by using kernel 2.1.62.
> Good, 2.1.62 continually crashed at my place:) It's only stable
> again with 2.1.64 (and my patch).
> > Questions: why does 2.1.62 work? (Machine has been runing 4 days now -
> > but not without some odd quirks, not sure if its because its 2.1.X or if
> > not all is well with 512M still, or both)
> I don't know why it works. But _if_ it works there's no reason
> not to use it.
> > can whatever be incorporated into next 2.0.X kernel?
> It will be in 2.2, but a lot of things just _can't_ be ported
> to 2.0 due to some fundamental differences in memory management.
> > some odd quirks observed:
> > -our model ("CORSIM" - crashes at a
> > weird spot now with "error closing file." This may be related to new C
> > libraries however.
> > -users complain of FTPed file corruptions. (they might at fault)
> > -System crashed while FTPing a large file.
> > -mouse skips all around the screen at random times (and clicks on
> > things at random) as if someone was writing into /dev/mouse.
> > -telnet from or to any machine which is not also linux is bizarre
> > (suspect termcap library need upgrading or something, w/ respect to new
> > C library maybe)
> It all sounds like memory fragmentation is biting you... If you've
> got some spare time you could try 2.1.64 with my patch (newest
> version). I think the problems should be (at least) alleviated quite
> a lot. If problems remain I could add some debugging messages to the
> mm part and analize the log with you. We could add an even more
> agressive anti-fragmentation strategy...
> > Provided one of our staff figures out quirk#1, the rest seems minor. I
> > am curious to know other's experiences with 2.1.62 as a "production"
> > kernel rather than test kernel. Please email me
> > ( any other oddities that have been observed.
> 2.1.6[34] are running very stable here. 2.1.62 didn't (mainly due
> to memory fragmentation??)
> > Our "big" machine is usually under heavy use but occasionally over a
> > weekend there is some idle time that I could use to test SMP kernels and
> > memory management of 256M or 512M. Let me know if this is convienent for
> > anyone.
> Well, it sounds like the perfect machine to test my patch (it's
> been thoroughly (sp) tested on smaller (<64M) machines, but never
> on such a big machine.
> I'd really apreciate it if you would like to stress-test my patch
> on your machine and report any odd things (high CPU usage for
> kswapd and vhand in particular) to me. Then I'll fix those things
> next week...
> You can find my patch on linux-mama and my homepage
> <>.
> success,
> Rik.
> ----------
> Send Linux memory-management wishes to me: I'm currently looking
> for something to hack...