Previous commit introduced a couple of problems in this area, for GEX44 hosts:
- the amount of memory calculated for a VM/slice was based on number of cores and assumed 2:1 ratio of cpus to cores for stanard-gpu VM. It should be 1:1 ratio
- the memory allocated was computed in the allocator but the memory freed was taken from VmSize structure, resulting in different values on GEX44 hosts.
This commit fixes those problems.
- It computes the amount of memory based on vcpus, just like VmSize
- The update to the host is based on request.memory_gib value when not creating a slice
- The update to the host is based on the amount allocated for a slice when using slices
Added multiple tests to cover all those cases.
14efc7c3b0