Integrating CXL Attached Memory in ArgoDSM

Sammanfattning

Software-based distributed shared memory (DSM) systems provide a unified address space across nodes but suffer from high latencies caused by remote memory accesses that can degrade performance when working sets exceed local Dynamic Random Access Memory (DRAM). Compute Express Link (CXL) introduces a new class of cache-coherent attached memory that offers greater capacity than local DRAM at lower latency than network-based remote access. This thesis designs, implements, and evaluates support for CXL attached memory within ArgoDSM. Two integration strategies are explored: a second-level page cache that absorbs clean pages evicted from Argo's primary cache, and a relocation of Argo's page-copy buffer to CXL. The second-level cache is implemented in three progressively refined variants, culminating in a design that maps cached pages directly into Argo's address space to bypass Argo's handler. Evaluation on a single-node CXL testbed under a synthetic-latency model shows that a CXL-backed second-level cache reduces execution time by up to 47 % on read-heavy workloads and 25 % on write-heavy workloads when the working set fits in the cache. The benefit is conditional on access pattern and working-set size, and the page-copy buffer relocation yields no measurable improvement. The results indicate that CXL attached memory has a defensible role as a secondary cache tier in software-based DSM, while leaving validation on a fully equipped multi-node cluster as future work.