hwloc - I/O topology

[tzcnt]

hwloc exposes some information about adjacency of I/O devices to NUMA nodes. https://dshcherb.github.io/2019/02/02/interpreting-pcie-device-to-cpu-locality-information.html

This could be used to:

• associate I/O executors with specific devices (network, storage, GPU)
• pin I/O executor threads to the directly attached NUMA node
• partition multiple I/O executors for different NUMA nodes / I/O devices

In tandem with explicit NUMA/NUCA partitioning each I/O executor could be associated with a specific CPU executor that runs on the same NUMA node. Thus we could parallelize workloads that involve both I/O and CPU without cross-node sharing.

[tzcnt]

As of v1.3, you can call .add_partition() on the asio executor to assign it to specific cores or NUMA nodes. Future I/O executors will have the same capability. So you can use this to do the pinning manually.

However, I don't have the hardware necessary to develop and test the querying of I/O locality to different CPUs (yet) so you'll need to manually choose the correct cores to pin it to.