The high demand for memory capacity in modern datacenters has led to multiple lines of innovation in memory expansion and disaggregation. One such effort is Compute eXpress Link (CXL)-based memory expansion, which has gained significant attention. To better leverage CXL memory, researchers have built several emulation and experimental platforms to study its behavior and characteristics. However, due to the lack of commercial hardware supporting CXL memory, the full picture of its capabilities may still be unclear to the community. In this work, we explore CXL memory's performance characterization on a state-of-the-art experimental platform. First, we study the basic performance characteristics of CXL memory using our proposed microbenchmark. Based on our observations and comparisons to standard DRAM connected to local and remote NUMA nodes, we also study the impact of CXL memory on end-to-end applications with different offloading and interleaving policies. Finally, we provide several guidelines for future programmers to realized the full potential of CXL memory
翻译:现代数据中心对于内存容量的高需求导致了多线创新,例如内存扩展和分离。其中基于Compute eXpress Link(CXL)的内存扩展已备受关注。为了更好地利用CXL内存,研究人员建立了几个仿真和实验平台,以研究其行为和特性。然而,由于缺乏商用硬件支持CXL内存,其全部性能还可能不清楚。在本研究中,我们探索了在最先进的实验平台上的CXL内存性能表现。首先,我们使用提出的微基准程序研究CXL内存的基本性能特征。根据我们的观察和与连接到本地和远程NUMA节点的标准DRAM的比较,我们还研究了CXL内存对多个端到端应用程序具有不同卸载和交错策略的影响。最后,我们提供了几个指导方针,以帮助未来程序员实现CXL内存的全部潜力。