| |||||||||||||
Time to thrash the 2.6 VM?Time to thrash the 2.6 VM?Posted Mar 4, 2004 19:39 UTC (Thu) by jwb (guest, #15467)Parent article: Time to thrash the 2.6 VM? I boggle when certain kernel developers claim that the current VM is useful on medium workloads. On my 8GB SMP machines, I often see userspace making no progress for minutes at a stretch, as seen in this chart: http://saturn5.com/~jwb/prime-starve.png The current VM can spin and spin doing nothing for a horribly long time. I'm glad AA is working on the problem. RvR, on the other hand, *is* the problem and I wish they would stop letting him screw around with the VM code. That's the way I see it from a sysadmin/user perspective, anyway.
(Log in to post comments)
Time to thrash the 2.6 VM? Posted Mar 4, 2004 19:58 UTC (Thu) by crimsun (subscriber, #13750) [Link] It's wrong to blame one person or any group for filibustering vm improvements. VM itself, as with any core OS component, is a very elusive bullseye: there is no best implementation for the average workload. 2.6 does a decent approximation. Sure, there's room for improvement - and the code is freely available for free modification and distribution.
Time to thrash the 2.6 VM? Posted Mar 4, 2004 20:45 UTC (Thu) by jwb (guest, #15467) [Link] I suppose it would be better to blame the process, but RvR's style of magically tweaking constants strikes me as unscientific. I'd be happier with an OSDL project which analyzed any proposed VM changes (or any kernel change) on a variety of workloads: kernel compiles on 128MB machines, interactive loads on 256MB machines, DVD burns on 1GB machines, web serving on 1 and 2GB machines, database loads on 1, 2, 4, 8, 16, and 32GB machines, and so forth.My enthusiasm for AA's changes stems from his proven history of having insight into how the kernel is really behaving, and his willingness to acknowledge the current brokenness.. RvR's 2.4 VM through 2.4.9 was a disaster. 2.4.10 was the climax, after which the kernel's behavior was so very much more reasonable. And I never forget that RvR invented the OOM killer, probably the worst decision for Linux operability ever made.
Time to thrash the 2.6 VM? Posted Mar 5, 2004 13:26 UTC (Fri) by Johbe (guest, #249) [Link] I'm having this problem right now. We run a squidproxy for about 1500 clients, a smp machine with approx 4 gigs of ram. Someone suggested I tried out Rik's rmap patch and that's what I'm doing right now.The problem I'm experiencing occationally is a complete lockup, kswapd eating *all* cpu on the machine, it becomes unusable, locked down to ssh and everything for 10-15 minutes then it starts running again as usual. I've been using the rmap patch for 3 days now, and so far, we haven't experienced this problem - it might show up, takes a while after a reboot until it hits. We'll see. But I still have some faith in it.
Time to thrash the 2.6 VM? Posted Mar 5, 2004 19:58 UTC (Fri) by jzbiciak (✭ supporter ✭, #5246) [Link] Just a quick question: Are these 8GB machines 32-bit machines running w/ highmem, or 64-bit machines running flat memory?Historically, it seems like highmem has always been a challenge to get to work well for Linux.
Time to thrash the 2.6 VM? Posted Mar 5, 2004 21:42 UTC (Fri) by jwb (guest, #15467) [Link] They are 64-bit machines running a 32-bit kernel with high memory. How's that for confusion?
|
|||||||||||||