Quelques commentaires (très favorables en ce qui concerne les changements majeurs apportés à la mémoire virtuelle par le magicien Andrea Arcangeli, le swap, les temps de réponse,...):
<<The VM has been totally rewritten by Andrea Arcangeli using his idea of 'classzone' balancing, and was merged in at 2.4.10-pre11.
I've been using 2.4.10-pre13 for a couple of days and... it's AMAZING! All your swap problems will disappear, and for me at least, the VM system is ~3x faster than any 2.4 kernel before it, adn faster than any 2.2 kernel as well by some distance.
Go get it whilst it's hot! No longer must we look enviously at FreeBSD!>>
________________________________
<< The VM is greatly improved. It uses less swap, and swap is released rather than just accumlating.>>
________________________________
Enfin, tout le monde n'est pas d'accord!
<< Well, AA contributed an enormous VM patch that basically changes the whole system. Apparently it has good effect for interactive uses like MP3 players and web browsing, but testers at HP labs say that the performance of the 2.4.10 VM is the worst of the (very bad already) 2.4.x series on their 4-8 GB machines with 30+ SCSI devices each. They make this conclusion based on NFS benchmarks.
On my machines, I've had tons of problems, and 2.4.10-preXX didn't make them go away. Until Linux drops the concept of memory overcommit, I'm afraid that the VM is going to continue to suck.>>
_____________________________
Je dirais même plus!
<< If your system is out of memory (totally out - no RAM, so Swap) then stuff will fail. That's true on any system
That is untrue. There is no need for any existing process to die in an OOM condition, although it may be necessary to prevent new processes from being spawned. The problem that the OOM killer tries to address is exhaustion of paging space by a rogue process, but that can be handled adequately by higher-level policy means (e.g. per-process limits, applying quota to paging space). Rejecting the policy-based approach in favor of warping the entire VM system's page-replacement strategy is IMO a mistake of the first magnitude. Contrary to what you imply, most other systems get by just fine without an OOM killer; many of them - but notably not including any flavor of NT - behave much more gracefully under memory pressure than Linux does.
Before you attempt to lecture me on how VM systems work, I should point out that I was at one time the go-to guy for a VM system more sophisticated than Linux's on one of the earlier SMP UNIXes (Encore's UMAX V). I do know how these things work, and that is why I make the claims that I do. The OOM killer does deserve to be killed.>>
________________________________
<< Yes. Yes, yes, yes. I can start UT on a 256-mb machine, play for an hour, come back out into X and still not have any swap used. It passes my stress-test ... we'll see about 3 days without a reboot or swapoff / swapon...>>
________________________________
<< There have been some persistent VM bugs for several versions (since about 2.4.4). 2.4.10 fixes them because Linus incorporated Andrea Arcangeli's VM patches. I'm running 2.4.10pre13aa1 and things are vastly improved. The "swap storms" of previous versions have completely gone away. 2.4.10 should be excellent.
If you still find swap problems, grab Andrea's latest patches (look in the people/andrea directory on the kernel mirrors). He's added some modified swap code the other day that's not in 2.4.10.>>
_____________________________
Zzzzooooooooooooommmmiiiiiiiinnnn'!!!!!!!!!
<< Current bandwidth utilization: 96.11 Mbit/s
That means:
1 kernel every 2-3 seconds
or
5 patches every second
or
96% of the bandwith from kernel.org is used>>
_____________________________
Desktop users may like the pre-emption patch
<<Those of you who use Linux as a desktop may be interested in the pre-emptible kernel patch for 2.4.10, available from here [tech9.net].
This patch allows the rescheduling of in-flight kernel syscalls if a higher-priority process than the process calling the syscalls becomes eligible to run.
What it means in practice for the typical desktop user is a major enhancement to interactive performance under Linux, especially when under heavy load. Your X pointer will never freeze with this patch. Using this patch, I have played skip-free mp3's whilst my system has had a loadavg of 20, and my KDE desktop was still usable. I could never hope to achieve this with ordinary Linux. It's a really impressive bit of work. Go try it out.
Of course, people with the need for proper real-time response out of Linux (musicians, for example) will love it even more... maximum latencies for me with this patch are under 4ms - again, very impressive.
It's slated for inclusion in the mainline kernel early in 2.5, but could do with lots of testing first... you know what to do.>>
_____________________________
Re:Desktop users may like the pre-emption patch
<< I've been running this patch on my desktop and notebook systems for a few days now, and my experiances have been very(!) positive. Desktop interaction, games, movies and music all interact a lot smoother (quite noticable on high loads, like compiles etc)
However as someone pointed out, the average throughput does suffer somewhat (last figure i heard was 4% decrease). This might not be the right thing for a server.
However if you primary use a system as a desktop system, go for it and try it out! it could use all the eyeballs it can get. >>
______________________________
<< Just downloaded it and built it. I just came off of a 2.4.5 kernel ... was used to an instant free memory drain -- 512Mb used to go to about 40Mb free in no time with X up. WOOHOO ... now showing 328Mb free with full KDE 2.1, xawtv, mozilla, and a few other thingies running - 60 processes in all. I think I'm gonna like this!!>>
