Investigating the Memory Extension and Disaggregated Memory Pooling System

Compute Express Link (CXL) enables high bandwidth, cache-coherency memory extension, memory pooling and sharing among multiple hosts, which increases the performance per CPU core and reduces the stranded memory in the datacenter. These features enable CXL to be applied in multiple cases including artificial intelligence (AI), in-memory database, and elastic cloud etc. This paper introduces a CXL memory extension and pooling system which includes a CXL extension card called Quartz, a memory extension box called Amber, a CXL switch box called Coral. In the system, there is a management engine that can control the CXL switch through sideband or in-band interface, enabling memory dynamic allocation. The testing results show that CXL can provide extra memory channel for extended memory bandwidth at a reasonable low memory access latency, and the memory tiering can further improve the system performance. The memory extension and pooling system can be applied for the artificial intelligence application to increase the training/inference performance, and can be applied for the in-memory database applications to increase the database query throughput.