Explore an innovative optimization technique for message-passing memory allocators in this 20-minute conference talk from ISMM 2024. Dive into the concept of BatchIt, a per-slab cache system designed to enhance performance in producer-consumer workloads. Learn how this optimization reduces contention on allocator synchronization mechanisms by batching remote deallocations destined for the same slab. Discover the implementation details for both mimalloc and snmalloc allocators, and examine experimental results showcasing over 20% performance improvements in certain scenarios. Gain insights into how BatchIt exploits naturally-arising locality of allocations to reduce atomic operations and cache misses, ultimately boosting efficiency in high-performance memory allocation systems.
BatchIt: Optimizing Message-Passing Allocators for Producer-Consumer Workloads
ACM SIGPLAN via YouTube
Overview
Syllabus
[ISMM24] BatchIt: Optimizing Message-Passing Allocators for Producer-Consumer Workloads
Taught by
ACM SIGPLAN