Rethinking Applications' Address Space with CXL Shared Memory Pools

Compute Express Link (CXL) is an emerging cache coherence interconnects standard that enables new forms of memory pooling and sharing in data centers.Today's data centers are built with many network-interconnected machines.While memory sharing could be implemented by software-managed coherence -like distributed shared memory (DSM), data centers prefer distributed computing as an alternative.Will the upcoming CXL 3.0 with hardware-managed coherence change the game?To start answering this question, in this paper we compare the strengths and weaknesses of hardware-managed coherence and software-managed coherence in the context of CXL 3.0, showing that neither is generally optimal.We propose an Adaptive Coherence Management strategy at the Operating System (OS) level that dynamically selects the most suitable coherence management mechanism at runtime, based on an application memory access pattern, and CXL memory access latency.The proposed Adaptive Coherence Management rethinks today's application's address spaces, introducing variable coherence management mechanisms for different parts of an address space, at runtime.We envision such parts of the address space being larger than a page but likely smaller than a Virtual Memory Area -this paper discusses our initial design.