Efficient and Scalable Synchronization via Generalized Cache Coherence

We explore the design of efficient and scalable synchronization for disaggregated shared memory. Porting existing synchronization primitives to such architectures results in poor performance scaling due to redundant inter-cache communications, exacerbated by high cache-coherence latency in disaggregated shared memory.

Driven by our insight that synchronization is a generalization of cache coherence in time and space, we argue for minimally extending existing cache coherence protocols to support synchronization primitives, thereby eliminating the redundant inter-cache communication inherent in layered synchronization. We propose a novel Generalized cache-Coherence Protocol (GCP) that realizes this insight by leveraging wait queues and variable-size cache lines directly at the cache-coherence layer for temporal and spatial generalization, respectively. We have verified GCP’s correctness using model checking. We present Soul, an end-to-end system implementation of GCP atop a disaggregated shared-memory platform. Soul supports popular lock APIs through a user-space library that offers improved performance without requiring any changes to application code. Our evaluation of Soul against state-of-the-art locks shows that it improves the performance of unmodified real-world applications at scale by 1–2 orders of magnitude while incurring <8% storage overhead.