No fixed latency for page migration?

[ch1y0q]

Is there an option to set a fixed latency for page migration?

It can be observed that page migration in mgpusim is modeled much slower than direct cache-line access. Is it due to more transferred data via RDMA (4KB vs. 64B) or simply due to many other penalties (TLB shootdown, CU pipeline flush, etc.)? Is there a way to manually set the overhead to a fixed value?

[syifan]

It is reasonable that migration is much slower than cache-line access. Both factors you described make a major difference.

Currently, since we model details of the migration process, there is no easy way to support fixed-latency migration. If you are interested, please consider discussing potential solutions here.