Linux.Ars features a detailed description of one of the most important parts of the newly-released Linux 2.6 kernel: the scheduler. The new scheduler features several improvements over that in 2.4; they will not only explain the improvements, but also describe how the scheduler works and why these improvements are important.
Inside the Linux 2.6 Scheduler
About The Author
Eugenia Loli
Ex-programmer, ex-editor in chief at OSNews.com, now a visual artist/filmmaker.
Follow me on Twitter @EugeniaLoli
Here is a slightly more technical version of the article (unedited)
http://67.163.90.104/~jim/scheduler.html
Well, I read both, and the one you posted is curtainly more in depth, although admitedly a little above my head. I think I get the basic concept though.
Anyway, I am mainly posting to ask how accurate the statement is of 2.6 efficiently scaling to 512 processors. I heard 2.4 could only handle maximum of 32, thus this seems very unbelievable, also, would you happen to have any data about performance of such a system? I have never seen a system with more then 8 processors, so this is kinda intregueing (sp) to me.
Anyway, I am mainly posting to ask how accurate the statement is of 2.6 efficiently scaling to 512 processors. I heard 2.4 could only handle maximum of 32, thus this seems very unbelievable, also, would you happen to have any data about performance of such a system? I have never seen a system with more then 8 processors, so this is kinda intregueing (sp) to me.
SGI supposedly exhibited machines that scaled 512 processors using Linux. I’m not sure if their patches are yet available in the major branches of Linux. And I cannot attest to their performance.
http://www.linuxplanet.com/linuxplanet/reports/5096/1/
Thanks for the link. The Ars Technica article really doesn’t describe how the Linux 2.6 scheduler works at all, but provides a conceptual overview of how schedulers work in general. For the real nitty gritty hit the link from the first post.
Some of the main contributors to this big advancement are companies like IBM. They want to allow to ship thier huge system with linux on them. Though I agree I havnt seen more then 16 processor systems available, but then again, I am not expert
Linux is not a hippie basement project anymore
Linux is not *just* a Hippie basement OS anymore. But as long as there are some hippies in basements hacking on the linux kernel, Linux will remain a Hippie Basement OS 🙂
> Linux will remain a Hippie Basement OS 🙂
which is… good, possibly!?
— “Anyway, I am mainly posting to ask how accurate the statement is of 2.6 efficiently scaling to 512 processors. I heard 2.4 could only handle maximum of 32, thus this seems very unbelievable…” —
There is deffinatly a huge improvment in scalability, but keep in mind that 2.4 actually scaled very poorly at that number of CPUs. Its not that 2.6 is so insanly great, its that 2.4 was so bad.
well people have mentioned the magical 512 number and yes sgi have a test system with nasa, from what i hear 128 is more reliable a figure for standard 2.6.0
still until we see some detail benchmarks we can’t really say
a nice set would be comparing the 512 Altrix box with a 512origin running IRIX
Everyone has been giving me shit because I didn’t know the ULE scheduler was O(1). I’m sorry. How does the ULE scheduler in FreeBSD compare to the Linux 2.6 one? Anyone have benchmarks?
The new Linux scheduler works exactly the same as the SCO Unix sceduler in benchmarks, lol.
http://bulk.fefe.de/scalability/
This might answer your question…
He wasnt using the ULE scheduler in that benchmarks.
ULE is only becoming a mainstream part of FreeBSD as of 5.3-CURRENT, and was not a part of that scalability test.
OT: Just for the hell of it when I get back to my own machines, I am going to run the tests* myself on FreeBSD 5.2, DragonFly (whatever the latest build is), a small LFS based Linux 2.6 distro and NetBSD -CURRENT.
* http://bulk.fefe.de/scalability/
Although I am mainly a FreeBSD user, I can’t shake this feeling that soon NetBSD will scale better than all of the OSs at least for a little while before FreeBSD or Linux catches up again. The progress they made during the first few public runs of those scalability benchmarks was nuts.