A comprehensive evaluation of spatial co-execution on GPUs using MPS and MIG technologies
Spatial sharing boosts GPU efficiency up to 30% but memory contention hurts MPS by 30%.
A new study from researchers at Universidad Complutense de Madrid provides a comprehensive evaluation of NVIDIA's Multi-Process Service (MPS) and Multi-Instance GPU (MIG) technologies for spatial co-execution on GPUs. The paper, published in The Journal of Supercomputing and available on arXiv, reveals a crucial trade-off between MPS's flexibility and MIG's isolation. In the most favorable scenarios, MPS improves performance by up to 30% and reduces energy consumption by about 20% when using its provisioning option to avoid resource monopolization. However, under memory contention, MPS suffers severe degradation, worsening performance by around 30%.
On the other hand, MIG's full hardware isolation resolves memory contention, leading to more consistent improvements. However, these gains are tempered by higher overhead, and its rigid partitioning scheme can degrade performance in certain cases. The study provides key insights for improving co-execution strategies according to job profiles, helping data centers and cloud providers optimize GPU utilization. The findings highlight that the choice between MPS and MIG depends on workload characteristics, with MPS excelling in flexible, low-contention environments and MIG offering reliable isolation for memory-intensive tasks.
- MPS improves GPU performance by up to 30% and reduces energy by 20% with provisioning
- MPS can degrade performance by 30% under memory contention due to lack of isolation
- MIG offers consistent improvements through full hardware isolation but has higher overhead and rigidity
Why It Matters
Helps data centers and cloud providers optimize GPU utilization and reduce costs by choosing the right sharing technology.