Explore a groundbreaking 21-minute conference talk from ISMM 2023 that introduces NUMAlloc, an innovative memory allocator designed for NUMA architecture. Delve into the challenges of existing allocators in supporting NUMA systems and discover how NUMAlloc's binding-based memory management approach addresses these issues. Learn about the "origin-aware memory management" technique for ensuring locality in memory operations and the "incremental sharing" method for balancing performance benefits with memory overhead when using transparent huge pages. Examine the extensive evaluation results showcasing NUMAlloc's superior performance, running 15.7% faster than the second-best allocator and 20.9% faster than the default Linux allocator. Gain insights into NUMAlloc's scalability up to 128 threads and its readiness for deployment in real-world applications.
Overview
Syllabus
[ISMM'23] NUMAlloc: A Faster NUMA Memory Allocator
Taught by
ACM SIGPLAN