Re: Questions on the scheduler
Oliver Herold wrote: OpenBSD isn't about performance, so it will be most of the time inferior. http://mail-index.netbsd.org/tech-kern/2007/09/28/0014.html Maybe this is of some help. But if compare it to Jeffs FreeBSD/Linux benches it looks rather strange to me. Yeah, that's the one I am talking about. He didn't provide any details of configuration, settings or tuning so it is not yet possible to understand what the graphs mean, if anything. Kris ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
Abdullah Ibn Hamad Al-Marri wrote: On 9/29/07, Kris Kennaway <[EMAIL PROTECTED]> wrote: Oliver Herold wrote: Are there any numbers or technical papers? Just out of curiosity. I ran a mysql benchmark against Dragonfly-current and FreeBSD 7 on an 8-core machine (one of the workloads that FreeBSD now performs very well at) and found 0 scaling on dragonfly. Their developers confirmed that the kernel is still entirely giant locked (as in FreeBSD 4) so no SMP performance benefits are possible. The email thread is here: http://leaf.dragonflybsd.org/mailarchive/users/2007-05/msg00134.html although the linked graph is offline. The FreeBSD curve was essentially this one (FreeBSD has improved further since then): http://people.freebsd.org/~kris/scaling/scaling.png with dragonfly a flat line at ~500 tps independent of load. Kris How does NetBSD, and OpenBSD scale when it comes to SMP comparing to FreeBSD 7.0? I was unable to boot NetBSD reliably on my hardware (the serial console only worked about 1/8 of the time) and did not even bother with OpenBSD because there is no reason to think they will be a contender for performance. One of the NetBSD developers recently posted a comparison on old 4*pentium 3 hardware, but their numbers are highly suspicious to me since they are way out of line with what I have measured on similar FreeBSD systems. I am waiting to hear back from him about it. Kris ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
OpenBSD isn't about performance, so it will be most of the time inferior. http://mail-index.netbsd.org/tech-kern/2007/09/28/0014.html Maybe this is of some help. But if compare it to Jeffs FreeBSD/Linux benches it looks rather strange to me. Cheers, Oliver On Sat, Sep 29, 2007 at 08:08:53PM +0300, Abdullah Ibn Hamad Al-Marri wrote: > On 9/29/07, Kris Kennaway <[EMAIL PROTECTED]> wrote: > > Oliver Herold wrote: > > > Are there any numbers or technical papers? Just out of curiosity. > > > > I ran a mysql benchmark against Dragonfly-current and FreeBSD 7 on an > > 8-core machine (one of the workloads that FreeBSD now performs very well > > at) and found 0 scaling on dragonfly. Their developers confirmed that > > the kernel is still entirely giant locked (as in FreeBSD 4) so no SMP > > performance benefits are possible. > > > > The email thread is here: > > > >http://leaf.dragonflybsd.org/mailarchive/users/2007-05/msg00134.html > > > > although the linked graph is offline. The FreeBSD curve was essentially > > this one (FreeBSD has improved further since then): > > > >http://people.freebsd.org/~kris/scaling/scaling.png > > > > with dragonfly a flat line at ~500 tps independent of load. > > > > Kris > > How does NetBSD, and OpenBSD scale when it comes to SMP comparing to > FreeBSD 7.0? > > > -- > Regards, > > -Abdullah Ibn Hamad Al-Marri > Arab Portal > http://www.WeArab.Net/ > ___ > freebsd-questions@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-questions > To unsubscribe, send any mail to "[EMAIL PROTECTED]" -- Surprise due today. Also the rent. ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
On 9/29/07, Kris Kennaway <[EMAIL PROTECTED]> wrote: > Oliver Herold wrote: > > Are there any numbers or technical papers? Just out of curiosity. > > I ran a mysql benchmark against Dragonfly-current and FreeBSD 7 on an > 8-core machine (one of the workloads that FreeBSD now performs very well > at) and found 0 scaling on dragonfly. Their developers confirmed that > the kernel is still entirely giant locked (as in FreeBSD 4) so no SMP > performance benefits are possible. > > The email thread is here: > >http://leaf.dragonflybsd.org/mailarchive/users/2007-05/msg00134.html > > although the linked graph is offline. The FreeBSD curve was essentially > this one (FreeBSD has improved further since then): > >http://people.freebsd.org/~kris/scaling/scaling.png > > with dragonfly a flat line at ~500 tps independent of load. > > Kris How does NetBSD, and OpenBSD scale when it comes to SMP comparing to FreeBSD 7.0? -- Regards, -Abdullah Ibn Hamad Al-Marri Arab Portal http://www.WeArab.Net/ ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
Thanks :-) Cheers, Oliver On Sat, Sep 29, 2007 at 06:45:20PM +0200, Kris Kennaway wrote: > Oliver Herold wrote: >> Are there any numbers or technical papers? Just out of curiosity. > > I ran a mysql benchmark against Dragonfly-current and FreeBSD 7 on an > 8-core machine (one of the workloads that FreeBSD now performs very well > at) and found 0 scaling on dragonfly. Their developers confirmed that the > kernel is still entirely giant locked (as in FreeBSD 4) so no SMP > performance benefits are possible. > > The email thread is here: > > http://leaf.dragonflybsd.org/mailarchive/users/2007-05/msg00134.html > > although the linked graph is offline. The FreeBSD curve was essentially > this one (FreeBSD has improved further since then): > > http://people.freebsd.org/~kris/scaling/scaling.png > > with dragonfly a flat line at ~500 tps independent of load. > > Kris > > ___ > freebsd-questions@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-questions > To unsubscribe, send any mail to > "[EMAIL PROTECTED]" -- Per buck you get more computing action with the small computer. -- R. W. Hamming ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
Oliver Herold wrote: Are there any numbers or technical papers? Just out of curiosity. I ran a mysql benchmark against Dragonfly-current and FreeBSD 7 on an 8-core machine (one of the workloads that FreeBSD now performs very well at) and found 0 scaling on dragonfly. Their developers confirmed that the kernel is still entirely giant locked (as in FreeBSD 4) so no SMP performance benefits are possible. The email thread is here: http://leaf.dragonflybsd.org/mailarchive/users/2007-05/msg00134.html although the linked graph is offline. The FreeBSD curve was essentially this one (FreeBSD has improved further since then): http://people.freebsd.org/~kris/scaling/scaling.png with dragonfly a flat line at ~500 tps independent of load. Kris ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
Are there any numbers or technical papers? Just out of curiosity. Cheers, Oliver On Sat, Sep 29, 2007 at 06:10:50PM +0200, Kris Kennaway wrote: > RW wrote: > >> The FreeBSD response was to make the kernel more SMP friendly with >> finer-grained locking, and to bring-in the ULE scheduler. Dragonfly BSD >> was a fork off 4.x by people who thought a more radical kernel rewrite >> was needed. Their kernel avoids a lot of the locking problems by using >> message queues. > > Just to clarify, that was the theory and intention behind Dragonfly, but in > practise they have yet to achieve it after 4 years and any benefits of > their ideas remain unproven. Basically they have achieved no performance > gains on SMP and have effectively abandoned working on it. > > Kris > ___ > freebsd-questions@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-questions > To unsubscribe, send any mail to > "[EMAIL PROTECTED]" -- Why did the Lord give us so much quickness of movement unless it was to avoid responsibility with? ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
RW wrote: The FreeBSD response was to make the kernel more SMP friendly with finer-grained locking, and to bring-in the ULE scheduler. Dragonfly BSD was a fork off 4.x by people who thought a more radical kernel rewrite was needed. Their kernel avoids a lot of the locking problems by using message queues. Just to clarify, that was the theory and intention behind Dragonfly, but in practise they have yet to achieve it after 4 years and any benefits of their ideas remain unproven. Basically they have achieved no performance gains on SMP and have effectively abandoned working on it. Kris ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
On Mon, 24 Sep 2007 10:23:40 -0400 "Jim Stapleton" <[EMAIL PROTECTED]> wrote: > I've heard a lot of winging about the FreeBSD scheduler from Linux > people, and even saw that is the reason for one fork off of FreeBSD. > > In my experience, I've gotten better performance out of FreeBSD on > single or multi-CPU systems than I have out of Linux or Windows (or > really any other system). > > Are these complains I hear of outdated, Probably. I think you are talking about SMP performance. A few years ago FreeBSD had good performance on single CPUs, but didn't scale very well onto machines with multiple CPUs, unlike Linux and Solaris. The kernel wasn't really designed to work this way and a lot of the code was protected by a single "Giant Lock". Companies like Yahoo and Hotmail (pre-Microsoft) tended to use FreeBSD where the load could be shared between many low-end machines. The FreeBSD response was to make the kernel more SMP friendly with finer-grained locking, and to bring-in the ULE scheduler. Dragonfly BSD was a fork off 4.x by people who thought a more radical kernel rewrite was needed. Their kernel avoids a lot of the locking problems by using message queues. As I understand it the locking problems were addressed in 5/6-current. There are still problems with the ULE scheduler in 6.x, but they have been fixed in 7-current, and things scale roughly as they should with multiple cores/cpus. ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Re: Questions on the scheduler
In response to "Jim Stapleton" <[EMAIL PROTECTED]>: > I've heard a lot of winging about the FreeBSD scheduler from Linux > people, and even saw that is the reason for one fork off of FreeBSD. > > In my experience, I've gotten better performance out of FreeBSD on > single or multi-CPU systems than I have out of Linux or Windows (or > really any other system). > > Are these complains I hear of outdated, am I hallucinationg, or is > there another answer? Although this is probably a better question for > -current, what is the state of the scheduler(s), and what would be > some good reading on the subject (specifically to BSD, and not just > schedulers in general)? Any discussion regarding such things is obsolete as quickly as it's written. Every OS I know if is constantly working to improve such things. Getting reliable, high-performance scheduling on modern SMP hardware is tough, but they all keep improving. Without a specific problem referencing a specific version, it's just idle chatter and useless for anything other than exercising your jaw between beers. If you have a specific performance problem, I highly recommend you file a PR with plenty of details. This is what happened with both MySQL and PostgreSQL and the result is that FreeBSD 7's ability to run those applications has improved dramatically. I doubt you're hallucinating, but without specifics, it's difficult to say what you're hearing. Lots of people think they can do benchmarking, but few (in my experience) are capable of legitimately doing a non-biased comparison that can really be trusted. I don't know where to point you for reading materials other than the code itself, and that's not something that's easily digested. As I said, writing high-quality schedulers is black magic, and the code reads that way. -- Bill Moran http://www.potentialtech.com ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"
Questions on the scheduler
I've heard a lot of winging about the FreeBSD scheduler from Linux people, and even saw that is the reason for one fork off of FreeBSD. In my experience, I've gotten better performance out of FreeBSD on single or multi-CPU systems than I have out of Linux or Windows (or really any other system). Are these complains I hear of outdated, am I hallucinationg, or is there another answer? Although this is probably a better question for -current, what is the state of the scheduler(s), and what would be some good reading on the subject (specifically to BSD, and not just schedulers in general)? Thanks, -Jim Stapleton ___ freebsd-questions@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-questions To unsubscribe, send any mail to "[EMAIL PROTECTED]"