12 ms·
As per sibling comment, this is about utilization efficiency and not breaking isolation (between MIG instances). The conclusion: > In this paper, we presented
by throw0101d 1y ago
As per sibling comment, this is about utilization efficiency and not breaking isolation (between MIG instances). The conclusion:
> In this paper, we presented MISO, a technique to leverage the MIG
functionality on NVIDIA A100 GPUs to dynamically partition GPU
resources among co-located jobs. MISO deploys a learning-based
method to quickly find the optimal MIG partition for a given job
mix running in MPS. MISO is evaluated using a variety of deep
learning workloads and achieves an average job completion time
that is lower than the unpartitioned GPU scheme by 49% and is
within 10% of the Oracle technique.