Closed ppebay closed 1 year ago
Adding the number of shared blocks as allowed rank QOI, with toy memory problem:
And with the "few" challenging problem:
Also added the number of "uprooted" shared blocks QOI (i.e. those that are present on a rank but where not initially homed there):
And in contrast, the number of "homed" blocks:
We can also look at the above in terms of ratio (or fraction) of homed blocks relative to the total number of blocks:
The same as above for the "toy" problem; unsurprisingly it was able to better retain original homing than for the more challenging problem:
So @lifflander is it the kind of QOI that would fit the bill and provide the kind of post-LB information that we would then want to summarize with a global statistic?
@PPP I like these QOIs.
For the challenging example visualized above, have you printed the details like in the very first comment at the top of the page to confirm that almost nothing is still at home at iteration 8/9? That seems pretty extreme.
If the visualization is correct, it seems like there is room for improvement in the load balancer as I would expect better performance by keeping the new groupings of objects that the load balancer found but changing on which rank the entire grouping will be processed so that there might be at least one object at home on more ranks.
In particular: