Leveraging Multi-Instance GPUs through moldable task scheduling

TL;DR

This paper introduces a novel scheduling algorithm, FAR, for multi-task execution on NVIDIA MIG GPUs, achieving near-optimal makespan with significant improvements over existing methods.

Contribution

The work presents a new 3-phase algorithm for moldable task scheduling on MIG GPUs, addressing reconfiguration costs and providing strong approximation guarantees.

Findings

FAR achieves a makespan within 1.22x of the optimal in real experiments.

The algorithm outperforms state-of-the-art scheduling methods.

Reconfiguration-aware scheduling significantly improves GPU utilization.

Abstract

NVIDIA MIG (Multi-Instance GPU) allows partitioning a physical GPU into multiple logical instances with fully-isolated resources, which can be dynamically reconfigured. This work highlights the untapped potential of MIG through moldable task scheduling with dynamic reconfigurations. Specifically, we propose a makespan minimization problem for multi-task execution under MIG constraints. Our profiling shows that assuming monotonicity in task work with respect to resources is not viable, as is usual in multicore scheduling. Relying on a state-of-the-art proposal that does not require such an assumption, we present FAR, a 3-phase algorithm to solve the problem. Phase 1 of FAR builds on a classical task moldability method, phase 2 combines Longest Processing Time First and List Scheduling with a novel repartitioning tree heuristic tailored to MIG constraints, and phase 3 employs local search via task moves and swaps. FAR schedules tasks in batches offline, concatenating their schedules on the fly in an improved way that favors resource reuse. Excluding reconfiguration costs, the List Scheduling proof shows an approximation factor of 7/4 on the NVIDIA A30 model. We adapt the technique to the particular constraints of an NVIDIA A100/H100 to obtain an approximation factor of 2. Including the reconfiguration cost, our real-world experiments reveal a makespan with respect to the optimum no worse than 1.22x for a well-known suite of benchmarks, and 1.10x for synthetic inputs inspired by real kernels. We obtain good experimental results for each batch of tasks, but also in the concatenation of batches, with large improvements over the state-of-the-art and proposals without GPU reconfiguration. Beyond the algorithm, the paper demonstrates the research potential of the MIG technology and suggests useful metrics, workload characterizations and evaluation techniques for future work in this field.