Re: [PATCH] Load balancing problem in 2.6.2-mm1

[Nick Piggin <piggin@cyberone.com.au>]

Martin J. Bligh wrote:

>>    Good stuff, I just gave the patch a spin and things seem a little
>>    calmer. However Im still seeing a lot of balancing going on within a
>>    node.
>>
>>This is a clearly recognizable edge case, so I'll try drawing this up on
>>some paper and see if I can suggest another patch.  There's no good reason
>>to move one lone process from a particular processor to another idle one.
>>
>>But it also approaches a question that's come up before:  if you have 2
>>tasks on processor A and 1 on processor B, do you move one from A to B?
>>One argument is that the two tasks on A will take twice as long as
>>the one on B if you do nothing.  But another says that bouncing a task
>>around can't correct the overall imbalance and so is wasteful.  I know
>>of benchmarks where both behaviors are considered important.  Thoughts?
>>
>
>It's the classic fairness vs throughput thing we've argued about before.
>Most workloads don't have that static a number of processes, but it 
>probably does need to do it if the imbalance is persistent ... but much
>more reluctantly than normal balancing. See the patch I sent out a bit
>earlier to test it - that may be *too* extreme in the other direction,
>but it should confirm what's going on, at least.
>
>

Yep. I've argued for fairness here, and that is presently what
we get. Between nodes the threshold should probably be higher
though.