Lightweight Memory Access Monitoring for Dynamic Data Placement in Tiered Memory Systems

With a growing number of memory technologies (e.g., PMEM, HBM, CXL memory), hardware characteristics and data access patterns on all available memory tiers must be monitored to make effective data placement decisions. Thus, memory management for optimal data placement across memory tiers becomes more CPU-and resource-demanding. Data placement can be either OS-controlled, e.g., through transparent virtual memory page promotion & demotion across memory tiers; or application-controlled, e.g., through explicit data allocation to specific memory tiers. However, OS-controlled data placement lacks awareness of workload resources, and application-controlled data placement results in additional overhead from managing access metrics and cost models. To facilitate dynamic data placement in tiered memory systems, modern OSs already provide built-in support for lightweight access monitoring to specific memory regions (e.g., through the Data Access MONitoring (DAMON) framework of the Linux Kernel). In this work, we investigate the applicability of the memory access monitoring framework DAMON in the context of database systems in tiered memory systems. To assess the applicability of DAMON as a core building block for data placement, we integrate its monitoring mechanism and memory management into Poseidon, a high-performance graph database system. In our experimental evaluation, we analyze the accuracy & overhead of DAMON through a set of microbenchmarks and by running end-to-end queries from a graph database benchmark against our DAMON-enabled Poseidon DBMS. Our initial results demonstrate the potential of DAMON as a core building block for workload-driven, dynamic data placement in tiered memory systems. Our experiments show a 3% overhead in execution time while achieving an accuracy of over 90% compared to actual access numbers.