[TransferEngine] add location argument to python register_memory methods

[thefacetakt]

Currently, there is no way to specify the location of a memory region in the Python API. As a result, every memory region is assigned to the wildcard (*) location, which makes explicitly specifying the topology (via MC_CUSTOM_TOPO_JSON) has no effect.

This PR adds support for an optional location argument in the Python API.

[stmatengss]

Could you add a unittest (testing different parameters) for the new register_memory API in transfer_engine_initiator_test.py?

[staryxchen]

Could you add a unittest (testing different parameters) for the new register_memory API in transfer_engine_initiator_test.py?

Hi @thefacetakt
I think your PR is very necessary. Could you add some tests based on the review comments? If you are not available at the moment, I would be happy to help.

[thefacetakt]

Hi @stmatengss

I've added a simple test which uses register_memory api with and without location argument.
Right now it basically tests that using this api does not break anything. It is challenging to test effect of providing location, as:

1. tcp transport ignores location
2. even with RDMA the effects are subtle and, i think, hard to test directly.

Did you have in mind that simple test, or do you have any ideas for more substantial test?

[stmatengss]

Hi @stmatengss

I've added a simple test which uses register_memory api with and without location argument. Right now it basically tests that using this api does not break anything. It is challenging to test effect of providing location, as:

1. tcp transport ignores location
2. even with RDMA the effects are subtle and, i think, hard to test directly.

Did you have in mind that simple test, or do you have any ideas for more substantial test?

A simple unit test is necessary (for both TCP and RDMA). You can try to add a unit test in memory_location_test.cpp. We aim to reduce parameters and ensure that any newly introduced parameter does not affect stability. THX!