GateKeeper: Transparent Placement of Big Data Objects in Hybrid Managed Heaps

Abstract

Popular big data frameworks running on top of managed runtimes must process datasets that typically outgrow DRAM capacity. Existing approaches extend the man- aged heap over slow but high-capacity media other than DRAM, such as NVMe SSDs. The state-of-the-art avoids costly garbage collection (GC) scans over the slow tier using two managed heaps, one in each tier. They use a DRAM (I/O) cache to offer fast access to objects in the slow tier. Nevertheless, these systems encounter signif- icant hurdles in determining which objects to migrate to the slow tier. One approach within existing systems relies on programming models or hints, demanding additional user effort and necessitating the rewriting of legacy frameworks, which is impractical in real-life deployments. An alternative category of systems identifies, transparently at runtime, the objects to move to the slow tier based on their object hotness. However, these solutions require costly code instrumentation, leading to a slowdown in application performance.