Primitives & BFS performance improvements

[seunghwak]

This PR includes multiple updates to cut peak memory usage in graph creation and improve performance of BFS on scale-free graphs.

• Add a bitmap for non-zero local degree vertices in the hypersparse region; this information can be used to quickly filter out locally zero degree vertices which don't need to be processed in multiple instances.
• Store (global-)degree offsets for vertices in the hypersparse region; this information can used to quickly identify the vertices with a certain global degree (e.g. for global degree 1 vertices, we can skip inter-GPU reduction as we know each vertex has only one neighbor).
• Skip kernel invocations in computing edge counts if the vertex list is empty.
• Add asynchronous functions to compute edge counts. This helps in preventing unnecessary serialization when we can process multiple such functions concurrently.
• Replace rmm::exec_policy with rmm::exec_policy_nosync in multiple places; the former enforces stream synchronization at the end. The latter does not.
• Enforce cache line alignment in NCCL communication in multiple places (NCCL communication performance is significantly affected by cache line alignment, often leading to 30-40% or more differences).
• For primitives working on a subset of vertices, broadcast a vertex list using a bitmap if the vertex frontier size is large. If the vertex frontier size is small (in case vertex_t is 8B and the local vertex partition range can fit into 4B), use vertex offsets instead of vertices to cut communication volume.
• Merge multiple host scalar communication function calls to a single one.
• Increase multi-stream concurrency in detail::extract_transform_e & detail::per_v_transform_reduce_e
• Multiple optimizations in template specialization (for update_major == true && reduce_op == any && key type is vertex && working on a subset of vertices) in detail::per_v_transform_reduce_e (this includes pre-processing vertices with non-zero local degrees; so we don't need to process such vertices using multiple GPUs, pre-filtering of zero local degree vertices, allreduce communication to reduce shuffle communication volumes, and special treatment of global degree 1 vertices, and so on).
• Multiple optimizations & specializations in detail::fill_edge_minor_property that works on a subset of vertices (this includes kernel fusion, specialization for bitmap properties including direct broadcast to the property buffer and special treatments for vertex partition boundaries, and so on).
• Added multiple optimizations & specializations in transform_reduce_v_frontier_outgoing_e (especially for reduce_op::any and to cut communication volumes and to filter out (key, value) pairs that won't contribute to the final results).
• Multiple low-level optimizations in direction optimizing BFS (including approximations in determining between bottom -up and top-down).
• Multiple optimizations to cut peak memory usage in graph creation.

[ChuckHastings]

Couple of minor comments.

[ChuckHastings]

Is this something we should look at updating universally, or is this still a tunable we should continue to evaluate?

[seunghwak]

We may individually tune this, but I think we may eventually create a (or few) separate file storing every tunable parameters.

[ChuckHastings]

Should this just be 2024?

[seunghwak]

Yes, and I am wondering whether we should keep this file or just delete it. We already have BFS tests, this is a bit redundant.