ROLP: Runtime Object Lifetime Profiling for Big Data Memory Management
Rodrigo Bruno,Duarte Patrício,José Simão,Luís Veiga,Paulo Ferreira
DOI: https://doi.org/10.48550/arXiv.1804.00702
2018-03-09
Distributed, Parallel, and Cluster Computing
Abstract:Low latency services such as credit-card fraud detection and website targeted advertisement rely on Big Data platforms (e.g., Lucene, Graphchi, Cassandra) which run on top of memory managed runtimes, such as the JVM. These platforms, however, suffer from unpredictable and unacceptably high pause times due to inadequate memory management decisions (e.g., allocating objects with very different lifetimes next to each other, resulting in memory fragmentation). This leads to long and frequent application pause times, breaking Service Level Agreements (SLAs). This problem has been previously identified and results show that current memory management techniques are ill-suited for applications that hold in memory massive amounts of middle to long-lived objects (which is the case for a wide spectrum of Big Data applications). Previous works try to reduce such application pauses by allocating objects off-heap or in special allocation regions/generations, thus alleviating the pressure on memory management. However, all these solutions require a combination of programmer effort and knowledge, source code access, or off-line profiling, with clear negative impact on programmer productivity and/or application performance. This paper presents ROLP, a runtime object lifetime profiling system. ROLP profiles application code at runtime in order to identify which allocation contexts create objects with middle to long lifetimes, given that such objects need to be handled differently (regarding short-lived ones). This profiling information greatly improves memory management decisions, leading to long tail latencies reduction of up to 51% for Lucene, 85% for GraphChi, and 60% for Cassandra, with negligible throughput and memory overhead. ROLP is implemented for the OpenJDK 8 HotSpot JVM and it does not require any programmer effort or source code access.