MEMPHIS: Holistic Lineage-based Reuse and Memory Management for Multi-backend ML Systems
Modern machine learning (ML) systems leverage multiple backends, including CPUs, GPUs, and distributed execution platforms like Apache Spark or Ray. Depending on workload and cluster characteristics, these systems typically compile an ML pipeline into hybrid plans of in-memory CPU, GPU, and distributed operations. Prior work found that exploratory data science processes exhibit a high degree of redundancy, and accordingly applied tailor-made techniques for reusing intermediates in specific backend scenarios. However, achieving efficient holistic reuse in multibackend MLsystems remains a challenge due to its tight coupling with other aspects such as memory management, data exchange, and operator scheduling.
In this paper, we introduce MEMPHIS, a principled framework for holistic, application-agnostic, multibackend reuse and memory management. MEMPHIS’s core component is a hierarchical lineage-based reuse cache, which acts as a unified abstraction and manages the reuse, recycling, exchange, and cache eviction across different backends. To address challenges of different backends such as lazy evaluation, asynchronous execution, memory allocation overheads, small available memory, and different interconnect bandwidths, we devise a suite of cache management policies. Moreover, we extend an optimizing ML system compiler by special operators for asynchronous exchange, workload-aware speculative cache management, and related operator ordering for concurrent execution. Our experiments across diverse ML tasks and pipelines show improvements up to 9.6x compared to state-of-the-art ML systems.
