irq_balance does not make sense with HT but single physical CPU

Hi @all,

this patch should disable irq_balance threat in case of only one installed physical cpu thats running in HyperThreading-mode (so reported as 2 cpus). I think it should make no sense to run irq_blanance in that special case - please correct me if i´m wrong.

Thanks,

Kai

diff -u -r -N linux-2.6.0-test10/arch/i386/kernel/io_apic.c linux-2.6.0-test10-kai/arch/i386/kernel/io_apic.c
@@ -627,6 +627,16 @@
 		irqbalance_disabled = 1;
 		return 0;
 	}
+
+#ifdef CONFIG_X86_HT
+	/* On Hyper-Threading CPUs - if only one physical installed
+	   balance does not make sense */
+	if (cpu_has_ht && smp_num_siblings == 2 && num_online_cpus() == 2) {
+		irqbalance_disabled = 1;
+		return 0;
+	}
+#endif
+
 	/*
 	 * Enable physical balance only if more than 1 physical processor
 	 * is present

+#ifdef CONFIG_X86_HT
+ /* On Hyper-Threading CPUs - if only one physical installed
+ balance does not make sense */
+ if (cpu_has_ht && smp_num_siblings == 2 && num_online_cpus() == 2) {
+ irqbalance_disabled = 1;
+ return 0;
+ }
+#endif

Further down, i believe the following would have the same effect;

	/*
	 * Enable physical balance only if more than 1 physical processor
	 * is present
	 */
	if (smp_num_siblings > 1 && !cpus_empty(tmp))
		physical_balance = 1;

tmp = cpu_online_map >> 2;

so we drop the first two processors (logical or physical) and only enable physical balance if there are still processors present in the map. Or are you observing something else?

Ok - inserted an printk(smp_num_siblings) to have a look into smp_num_siblings at that point.

It says : smp_num_siblings = 2

But anyways if physical_balance is set to 1 that won´t prevent anything from running through/sleeping in the kernel_thread-loop. The kernel_thread(balance_irq ...) later on will be started/will run not matter what physical_balance says.

Do there exist any cases where smp_siblings are created without HyperThreading ? (As far as I remember it´s only incremented/used on i386 hyperthreaded architectures - but not 100% sure)

-> At least the if has to look like :

...
if (smp_num_siblings > 2 && !cpus_empty(tm))

    physical_balance = 1;

...

(if - like in my case - smp_num_siblings == 2 on a single P4 system)

That one still does not solve the case in which the kernel_thread is not needed and only eats resources.
Of course - maybe the whole thing can be merged together. Not sure about that. But source will become more complex in that case - k.i.s.s. may be the better approach.

Thanks,

Kai

--------------------------------------------
Kai Bankett
Network Engineering

AOL Deutschland GmbH & Co. KG

Millerntorplatz 1
20359 Hamburg
Tel.:  +49 40 36159 - 7559
Fax.:  +49 40 36159 - 7510
Mobil: +49 172 2353870
eMail to  [email protected]
AIM:      kbankett
--------------------------------------------

Yes that only stops balancing across physical packages when there are none. But there might be a performance improvement for light (cache footprint wise) high frequency interrupt handling which stays affined to one logical processor.

This is all i386 specific code so we don't have to care about other architectures in here.

smp_num_siblings won't be greater than 2 with current i386 processors, it's not a total sibling count, but a per physical package count.

Does it makes no sense?

If you had two tasks both running continuously, one per sibling, then not IRQ balancing will penalise the task on the sibling which is getting most interrupts.

-- Jamie

Yes - did some more reading on HT and how the cpu looks like in that case ... I was thinking that interrupt lines - while in HT mode - aren´t available independently to both "parts" of the cpu. That was my fault - of course.

So just dismiss the patch - sorry.

Thanks,

Kai