Abstract

Memory connected via Compute Express Link (CXL) presents a new opportunity for building efficient and scalable databases. A CXL Pod is a small number of independent hosts connected via CXL to shared memory (termed CXL memory). A small part of the CXL memory is hardware-cache-coherent across hosts. This paper proposes Pasha, a new database architecture for the CXL Pod, that aims to efficiently leverage the fast synchronization enabled by the hardware-cachecoherence of CXL memory and the low latency and high bandwidth of local DRAM in each host. To achieve this goal, Pasha divides and stores data into partitions, with each partition owned by a single host, and a shared region that every host has access to. Then, Pasha uses CXL memory to efficiently synchronize concurrent accesses to the shared region across hosts and stores each partition of data in local DRAM to leverage its low latency without crosshost synchronization overhead. In addition, Pasha ensures that each host completes its transactions by directly reading or writing data in either its own partition or the shared region, circumventing multi-host transactions and two-phase commit (2PC). Pasha obtains the best of partition-based distributed databases and single-node shared-memory databases; preliminary results show it outperforms both architectures.