Looking into the Peak Memory Consumption of Epoch-Based Reclamation in Scalable in-Memory Database Systems

Deferred memory reclamation is an essential mechanism of scalable in-memory database management systems (DBMSs) that releases stale objects asynchronously to free operations. Modern scalable in-memory DBMSs commonly employ a deferred reclamation mechanism named epoch-based reclamation (EBR). However, no existing research has studied the EBR’s trade-off between performance improvements and memory consumption; its peak memory consumption makes capacity planning difficult and sometimes causes disruptive performance degradation. We argue that gracefully controlling the peak memory usage is a key to achieving stable throughput and latency of scalable EBR-based in-memory DBMSs. This paper conducts a quantitative analysis and evaluation of a representative EBR-based DBMS, Silo, from the viewpoint of memory management. Our evaluation reveals that the integration of conventional solutions fails to achieve stable performance with lower memory utilization, and Glasstree-based Silo achieves a 20% higher throughput, latencies characterized by an 81% lower standard deviation, and 34% lower peak memory usage than Masstree-based Silo even under read-majority workloads.