Re: Processor affinity planned for 2.5?
Wasnt my idea ... I´ll forward this to the guy who asked ... On 128 CPU machines affinity to processor groups may be a thing to consider but not on 2 or 4 way SMPs ... Anyway, I appreciate that he asked on linux-SMP and not here ... less traffic in here ... Markus Jakob Østergaard wrote: > > On Wed, Oct 18, 2000 at 01:13:48PM +0200, Markus Pfeiffer wrote: > > HI all > > > > I received and discussed Processor-affinity on the SMP-Kernel-list and I > > only wanted to ask if there is something in preparation or if someone > > (especially Linus :)) is against this ... I was asked this question > > beacause someone wanted to do some Matrix-Operations on his SMP machine > > and wnated to speed them up by trying affinity ... > > I don't know the exact answer to your question, but I can tell you that there > has been patches floating around for a long time (somehwere) to provide > processor affinity for processes in the Linux kernel. > > However, the patches has never been merged because: > *) No matter how self-contained and simple and elegant they may be, > they complicate the scheduler (maybe not much, but more than not) > *) There is no performance gain what so ever, expcept for perhaps some > very specially fabricated benchmark in some very specially fabricated > test environment. In the best case, you will not get a slowdown from > the explicit affinity. > > Think about it: It's the scheduler's job to distribute the threads of > execution reasonably among the available processors. The scheduler knows the > state of the system and *you* don't when you put in your static affinity > decisions. You are bound to lose. > > Yes, I know that big-iron IRIX boxes have explicit processor affinity available > to user-space. They also have gang-scheduling, which is more complicated and > maybe more likely to actually give you a speedup than explicit affinity. IRIX > has a lot of things, and let's not get into *that* one again ;) > > If you want to achieve anything in life, keep it simple. Be happy you have a > scheduler that on any day can do a better job than you at scheduling the > threads, and get on with your work :) > > Cheers, > -- > > : [EMAIL PROTECTED] : And I see the elder races, : > :.: putrid forms of man: > : Jakob Østergaard : See him rise and claim the earth, : > :OZ9ABN : his downfall is at hand. : > :.:{Konkhra}...: > - > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in > the body of a message to [EMAIL PROTECTED] > Please read the FAQ at http://www.tux.org/lkml/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Processor affinity planned for 2.5?
On Wed, Oct 18, 2000 at 01:13:48PM +0200, Markus Pfeiffer wrote: > HI all > > I received and discussed Processor-affinity on the SMP-Kernel-list and I > only wanted to ask if there is something in preparation or if someone > (especially Linus :)) is against this ... I was asked this question > beacause someone wanted to do some Matrix-Operations on his SMP machine > and wnated to speed them up by trying affinity ... I don't know the exact answer to your question, but I can tell you that there has been patches floating around for a long time (somehwere) to provide processor affinity for processes in the Linux kernel. However, the patches has never been merged because: *) No matter how self-contained and simple and elegant they may be, they complicate the scheduler (maybe not much, but more than not) *) There is no performance gain what so ever, expcept for perhaps some very specially fabricated benchmark in some very specially fabricated test environment. In the best case, you will not get a slowdown from the explicit affinity. Think about it: It's the scheduler's job to distribute the threads of execution reasonably among the available processors. The scheduler knows the state of the system and *you* don't when you put in your static affinity decisions. You are bound to lose. Yes, I know that big-iron IRIX boxes have explicit processor affinity available to user-space. They also have gang-scheduling, which is more complicated and maybe more likely to actually give you a speedup than explicit affinity. IRIX has a lot of things, and let's not get into *that* one again ;) If you want to achieve anything in life, keep it simple. Be happy you have a scheduler that on any day can do a better job than you at scheduling the threads, and get on with your work :) Cheers, -- : [EMAIL PROTECTED] : And I see the elder races, : :.: putrid forms of man: : Jakob Østergaard : See him rise and claim the earth, : :OZ9ABN : his downfall is at hand. : :.:{Konkhra}...: - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Processor affinity planned for 2.5?
On Wed, Oct 18, 2000 at 01:13:48PM +0200, Markus Pfeiffer wrote: HI all I received and discussed Processor-affinity on the SMP-Kernel-list and I only wanted to ask if there is something in preparation or if someone (especially Linus :)) is against this ... I was asked this question beacause someone wanted to do some Matrix-Operations on his SMP machine and wnated to speed them up by trying affinity ... I don't know the exact answer to your question, but I can tell you that there has been patches floating around for a long time (somehwere) to provide processor affinity for processes in the Linux kernel. However, the patches has never been merged because: *) No matter how self-contained and simple and elegant they may be, they complicate the scheduler (maybe not much, but more than not) *) There is no performance gain what so ever, expcept for perhaps some very specially fabricated benchmark in some very specially fabricated test environment. In the best case, you will not get a slowdown from the explicit affinity. Think about it: It's the scheduler's job to distribute the threads of execution reasonably among the available processors. The scheduler knows the state of the system and *you* don't when you put in your static affinity decisions. You are bound to lose. Yes, I know that big-iron IRIX boxes have explicit processor affinity available to user-space. They also have gang-scheduling, which is more complicated and maybe more likely to actually give you a speedup than explicit affinity. IRIX has a lot of things, and let's not get into *that* one again ;) If you want to achieve anything in life, keep it simple. Be happy you have a scheduler that on any day can do a better job than you at scheduling the threads, and get on with your work :) Cheers, -- : [EMAIL PROTECTED] : And I see the elder races, : :.: putrid forms of man: : Jakob Østergaard : See him rise and claim the earth, : :OZ9ABN : his downfall is at hand. : :.:{Konkhra}...: - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/
Re: Processor affinity planned for 2.5?
Wasnt my idea ... I´ll forward this to the guy who asked ... On 128 CPU machines affinity to processor groups may be a thing to consider but not on 2 or 4 way SMPs ... Anyway, I appreciate that he asked on linux-SMP and not here ... less traffic in here ... Markus Jakob Østergaard wrote: On Wed, Oct 18, 2000 at 01:13:48PM +0200, Markus Pfeiffer wrote: HI all I received and discussed Processor-affinity on the SMP-Kernel-list and I only wanted to ask if there is something in preparation or if someone (especially Linus :)) is against this ... I was asked this question beacause someone wanted to do some Matrix-Operations on his SMP machine and wnated to speed them up by trying affinity ... I don't know the exact answer to your question, but I can tell you that there has been patches floating around for a long time (somehwere) to provide processor affinity for processes in the Linux kernel. However, the patches has never been merged because: *) No matter how self-contained and simple and elegant they may be, they complicate the scheduler (maybe not much, but more than not) *) There is no performance gain what so ever, expcept for perhaps some very specially fabricated benchmark in some very specially fabricated test environment. In the best case, you will not get a slowdown from the explicit affinity. Think about it: It's the scheduler's job to distribute the threads of execution reasonably among the available processors. The scheduler knows the state of the system and *you* don't when you put in your static affinity decisions. You are bound to lose. Yes, I know that big-iron IRIX boxes have explicit processor affinity available to user-space. They also have gang-scheduling, which is more complicated and maybe more likely to actually give you a speedup than explicit affinity. IRIX has a lot of things, and let's not get into *that* one again ;) If you want to achieve anything in life, keep it simple. Be happy you have a scheduler that on any day can do a better job than you at scheduling the threads, and get on with your work :) Cheers, -- : [EMAIL PROTECTED] : And I see the elder races, : :.: putrid forms of man: : Jakob Østergaard : See him rise and claim the earth, : :OZ9ABN : his downfall is at hand. : :.:{Konkhra}...: - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/ - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] Please read the FAQ at http://www.tux.org/lkml/