[FEA]MG Pagerank

[afender]

The goal is to start from the current implementation (one process running on multiple GPUs) and transform it to run with one process per GPU.

The current version was designed to take advantage of DGX architectures. The new solution should work on cloud-based environments and non-nvlink systems.

We should bring OPG to cuGraph though generalizing and sharing the comm libs with cuML. That involves moving a generic version of cuML's handle and device allocator to RAFT and cuML's communicator to a separate comm lib. The comm part can't be in RAFT as it is designed to allow the user to swap out the implementation at runtime (so it is not headers only). We will then add these new dependencies in cuGraph and build OPG analytics on top of it.

C++/CUDA
We need to move from shared memory, multithreaded code to multi-process code at the CPP/CUDA level.

• COO2CSR
• Degree
• Pagerank setup
• Power iteratinon
• CAPI
• Tests

Dask/Python
Need to move from the IPC based solution to the recent solution deployed in cuML.

• python API backend upgrade
• Tests

Update Doc

Verify Notebooks