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

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

Rick Lindsley wrote:

>    We also shouldn't need load_diff, because if (avg_load <= this_load)
>    then imbalance will be zero, so I'll fix that up.
>
>Are we sure imbalance will be zero?  I think we still need that.  We can
>turn it into a single C statement if we want to be clever but what you
>save in temporary variable we'd better replace in comments to make the
>cleverness plain.  We need there to be no "negative" numbers in the
>min() statement in case max-avg is non-zero but avg-this is "negative".
>
>Imagine loads of
>
>    cpu0	0
>    cpu1	0
>    cpu2	3 
>    cpu3	2
>    cpu4	0
>
>and we're running on cpu3.
>
>    max_load=3
>    avg_load=1
>    this_load=2
>
>min(max-avg, avg-this) will be min(3-1,1-2) or two, and we'll choose to
>try to pull two to cpu3 instead of just leaving it alone which is the
>right thing to do.
>
>

No you definitely still need the test... this is what I mean:

       if (avg_load > this_load)
                *imbalance = min(max_load - avg_load, avg_load - this_load);
        else
                *imbalance = 0;