PaCaR: Improved Buffered I/O Locality on NUMA Systems with Page Cache Replication

Modern systems featuring multiple sockets suffer from the non-uniformity of performance of their memory accesses. Accessing memory of a remote node can lead to doubled latencies and halved bandwidth. This is particularly critical for I/O-intensive applications relying on the Linux page cache, where locality is paramount. Existing solutions migrating threads or memory across nodes fail to address the issue with highly parallel workloads, where a small working set is constantly accessed by multiple threads across nodes.

We present PaCaR, a novel page cache replication mechanism that enhances locality by transparently replicating cached pages across Non-Uniform Memory Access (NUMA) nodes. PaCaR ensures write consistency, adapts to memory pressure, and operates seamlessly without requiring application modifications. Evaluations on a dual socket NUMA server demonstrate up to 1.4x performance improvements in synthetic workloads and up to 25% gains in real-world workloads, showcasing the effectiveness of PaCaR in optimizing NUMA-aware buffered I/O operations.