From: Andrea A. <an...@su...> - 2002-07-24 12:40:16
|
On Wed, Jul 24, 2002 at 08:25:25AM +1000, Anton Blanchard wrote: > > > BTW, could somebody re-run the benchmarks that expectedly shows the bad > > regression of the original irq-balance patch and compare it with my new > > irq-balance algorithm that I partly rewrote while merging it in my > > tree? As far I could see (before ever hearing about any benchmark) the > > previous irqbalance was mainly a beauty-hack to make P4 smp look like a > > PIII smp in /proc/interrupts, some of the very obvious bugs: the > > uninitialized idle_timestamp, the unconditional overwriting of the > > ioapic settings even if the routing decision didn't change, the overkill > > frequency of the rounting change, the icache trashing due superflous > > routing migrations. > > The unconditional overwriting and high frequency of updates were the two > things I found when porting this to ppc64. Its especially bad now that > HZ = 1000, it should not be doing an update every time jiffies changes. of course while fixing it I also made the update correctly in function of HZ, so it doesn't matter what HZ it is. But really the only problems weren't the frequent updates and unconditional overwriting, the thing keeps bouncing if it doesn't find any better routing and it has the other issues mentioned above, all should be fixed in my rewrote version in -aa. Andrea |