25646, "jabraham17", "[Feature Request]: over-subscription for GPUs", "2024-07-25T23:02:03Z"
Today, Chapel allows you to oversubscribe both tasks and locales. This allows users to emulate more cores or nodes. It would be nice to also allow over-subscription with GPUs. We can already kinda do this today when emulating a GPU with CHPL_GPU=cpu using CHPL_RT_NUM_GPUS_PER_LOCALE.
It would be nice to wire this up to work with real gpus as well. Right now it is explicitly disallowed at runtime, but it might be nice to be able to simulate multiple gpus on a workstation that only has one gpu. This would allow users developing multi-gpu codes on a desktop with only one gpu to test their code with simulated/over subscribed gpus before moving to a larger system.