Microbenchmarking of Sampling/Loading APIs #290
Description
We want to better understand the runtime and memory usage of each component of the sampling/loading API in cuGraph-PyG to identify areas for improvement. In particular, memory usage is a good target.
HeterogeneousSamplerReader.__decode_coo is an excellent starting point - this function probably does the most significant amount of work per batch, and we've seen performance degradation with very small batch sizes (8, 16, 32) that can probably be blamed on how this function handles the bookkeeping per batch.