https://bugzilla.novell.com/show_bug.cgi?id=342593 Cyril Hrubis changed: What |Removed |Added ---------------------------------------------------------------------------- AssignedTo|bnc-team-screening@forge.provo.novell.com |bwalle@novell.com -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.

https://bugzilla.novell.com/show_bug.cgi?id=342593#c2 --- Comment #2 from Dieter Jurzitza 2007-11-19 15:05:37 MST --- Some mainboards are sensitive to distributing specific interrupts between different CPUs. More specifically this is the case for - at least - the Tyan Tiger and Tyan Thunder mainboards with Athlon CPU. At least those boards react allergic if you freely switch the interrupts between CPU0 and CPU1 - don't ask me why. It is a known phenomen, probably related to the AMD MP chipset. I have heard about other mainboards with similar issues, however, I can only discuss things I know about. So, IRQ-Balancer per default distributes interrupts on request, regardless whether your system is sensitive to it or not. On the mainboards above this means crashes, instability and what have you. There are options to suppress this behaviour, however, irq_balancer is not easy to configure in this regard. You would have to manually edit /etc/init.d/irq_balancer and insert environment variables to take influence. The IRQ-balancing benefit may be real on servers running heavy load. I can hardly imagine that this is the usecase for the vast majority of openSUSE users / customers. I'd guess: Who does not know about this option would not really care. The improvement in performance under normal home usage conditions is hardly perceivable. Who does know about it (and hence what he does) can activate by himself (and on his own risk). Alternatively a script allowing to share specific interrupts to specific CPUs would be (IMHO) more convenient to experienced users as it would provide the freedom to assign intentionally. (By the way: I could provide a rpm-package providing exactly this functionality. But this is not the point) Just my 2 cents here: - performance loss: small - risc reduction: significant hence: remove the kernel dependency and leave irq_balancer an option for those who know. The main target IMHO should be reliability, not the last little bit of performance. But again: my 2 cents here only. -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.

https://bugzilla.novell.com/show_bug.cgi?id=342593#c4 Bernhard Walle changed: What |Removed |Added ---------------------------------------------------------------------------- Status|NEW |NEEDINFO Info Provider| |ak@novell.com --- Comment #4 from Bernhard Walle 2007-11-26 07:41:25 MST --- Andi, I'd like to hear your opinion on that issue since you're the person that probably knows most about that. Maybe it also would be possible to blacklist some chipsets that are known to have problems. In general, if I turn it off by default I get LOTS of reports why it's turned off. ;-( -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.

https://bugzilla.novell.com/show_bug.cgi?id=342593#c6 Andi Kleen changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |ak@novell.com Status|NEEDINFO |NEW Info Provider|ak@novell.com | --- Comment #6 from Andi Kleen 2007-11-28 10:41:48 MST --- I'm not sure what the original reporter means with "difficult to configure" irq_balancer normally doesn't need any configuration in normal operation. Personally I think it's useless on single socket systems though, but on multi socket it's a good idea. -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.

https://bugzilla.novell.com/show_bug.cgi?id=342593#c8 Bernhard Walle changed: What |Removed |Added ---------------------------------------------------------------------------- Status|NEW |RESOLVED Resolution| |WORKSFORME --- Comment #8 from Bernhard Walle 2007-11-29 05:07:35 MST --- Ok, so we will not change the default. Dieter, can you please open a Bugzilla if you have problems with concrete hardware? -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.

https://bugzilla.novell.com/show_bug.cgi?id=342593 User bwalle@novell.com added comment https://bugzilla.novell.com/show_bug.cgi?id=342593#c10 Bernhard Walle changed: What |Removed |Added ---------------------------------------------------------------------------- Status|RESOLVED |REOPENED Resolution|WORKSFORME | --- Comment #10 from Bernhard Walle 2007-12-05 14:44:24 MST --- (In reply to comment #9 from Dieter Jurzitza)

Just for my understanding - I should open up a bug report for Tyan Tiger & Tyan Thunder mainboards with Athlon MP CPUs because they're sensitive against shared interrupts?

Can you send a dmidecode output for one of such a mainboard?

Seriously, can anyone give numbers with regard to the performance improvement on a home user system sharing interrupts against non sharing interrupts?

I also think that there's not performance improvement. However, in openSUSE 10.2 we had the IRQ balancer disabled for systems with one CPU but multiple cores. The result was that I got lots of bug reports about this. I think it's hard to give numbers and honestly I don't have time now to do measurements in that area. If there are mainboards with problems I add some blacklisting in the script based on dmidecode.

What regards configuration: you _can_ configure irq_balancer via environment variables, telling it to leave certain irq's "as is". But this is not possible in the openSUSE implementation as there are no corresponding options in irq_balancer startup script today.

I can add a sysconfig file for this, no problem.

Even if you wanted to make a fix assignment for certain interrupts to a certain CPU there is no way in the current package. I once promoted a scriptlett doing the same that irq_balancer does but with configuration options through yast.

Actually I see no need for a binary here, it is enough to have a set of appropriate

That's wrong. The binary does much more, please read http://www.irqbalance.org/. If you think that's not enough configurable, would it be possible for you to convince the upstream maintainer of that package? I wrote him once an email with a patch about the Makefile and never got a reponse from him. -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.

https://bugzilla.novell.com/show_bug.cgi?id=342593 User bwalle@novell.com added comment https://bugzilla.novell.com/show_bug.cgi?id=342593#c12 Bernhard Walle changed: What |Removed |Added ---------------------------------------------------------------------------- Status|REOPENED |RESOLVED Resolution| |FIXED --- Comment #12 from Bernhard Walle 2007-12-11 14:52:44 MST --- Added /etc/sysconfig/irqbalance in Factory. Please open a bugreport against concrete hardware and against kernel for the problems you still have! -- Configure bugmail: https://bugzilla.novell.com/userprefs.cgi?tab=email ------- You are receiving this mail because: ------- You are on the CC list for the bug.