Date: Fri, 27 Oct 2006 15:37:16 -0700 From: Julian Elischer <julian@elischer.org> To: Matthew Dillon <dillon@apollo.backplane.com> Cc: Paul Allen <nospam@ugcs.caltech.edu>, current@freebsd.org Subject: Re: Comments on the KSE option Message-ID: <45428A1C.3020902@elischer.org> In-Reply-To: <200610272044.k9RKijTe045236@apollo.backplane.com> References: <45425D92.8060205@elischer.org> <20061027201838.GH30707@riyal.ugcs.caltech.edu> <200610272044.k9RKijTe045236@apollo.backplane.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Matthew Dillon wrote: > I think the real issue here is that it is fairly difficult... probably > close to impossible in fact, to write a general purpose scheduler in > the kernel that can handle the types of extreme cases that can occur > when the kernel is made responsible for managing a user program's threads. > > A better approach would be to make the kernel responsible for scheduling > cpu slots for programs, a far more manageable number, and if a program > wants to have thousands of threads on a 4-cpu system it (i.e. libthr) > should then have the responsibility of telling the kernel which > threads to pop into those slots at any given moment. hey, wait, that's what the M:N library does! > > So, for example, if a machine has 4 cpus the kernel has 4 scheduling > slots it can fill. The kernel can apportion those slots with its current > scheduler technology. But if there is a program running on the system > that has thousands of threads, then why in the world would you want to > try to make the kernel scheduler deal with all those threads at once > when it only has 4 cpus to assign them to anyhow? So what you would > have instead would be the kernel saying to the program 'ok, I have 3 > slots available for you at the moment' and make the program responsible > for telling the kernel which threads to run in those 3 slots. And > then a little while later the kernel might say 'I have 4 slots now', > or 'now I only have 2 slots available', etc etc. > > This would then be a far more solvable problem for the kernel scheduler. > > If you as the user then want the kernel to give the program with > thousands of threads more of the available cpu, it simply becomes a > matter of running the program at a lower NICE value. > > -Matt
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?45428A1C.3020902>