Closed ppebay closed 1 year ago
When load-balancing the "toy memory problem":
[lbsStatistics] Descriptive statistics of initial rank loads:
[lbsStatistics] cardinality: 4 sum: 350 imbalance: 1.17143
[lbsStatistics] minimum: 20 mean: 87.5 maximum: 190
in 4 iterations leading to:
[lbsStatistics] Descriptive statistics of final rank loads:
[lbsStatistics] cardinality: 4 sum: 350 imbalance: 0.0857143
[lbsStatistics] minimum: 80 mean: 87.5 maximum: 95
we see that the shared memory properties went from a perfectly balanced situation:
[lbsStatistics] Descriptive statistics of final rank shared memory:
[lbsStatistics] cardinality: 4 sum: 1.92e+10 imbalance: 0
[lbsStatistics] minimum: 4.8e+09 mean: 4.8e+09 maximum: 4.8e+09
to the following:
[lbsStatistics] Descriptive statistics of final rank shared memory:
[lbsStatistics] cardinality: 4 sum: 1.76e+10 imbalance: 0.0909091
[lbsStatistics] minimum: 3.2e+09 mean: 4.4e+09 maximum: 4.8e+09
i.e., shared memory decreased. These results might be numerically incorrect (and probably are) but show that the implementation now allows for modification of shared memory properties as objects are transferred.
I suggest that numerical validation be made in a separate PR and that we review this one functionally before.
@lifflander @nlslatt
@lifflander @nlslatt I overhauled the CI; all passing now.
Ready for review when you get a chance.
Thanks!
Please @tlamonthezie confirm whether this breaks the CI/CD or not
Related to #314