SWAT: A Programmable, In-Memory, Distributed, High-Performance Computing Platform

The field of data analytics is currently going through a renaissance as a result of ever-increasing dataset sizes, the value of the models that can be trained from those datasets, and a surge in flexible, distributed programming models. In particular, the Apache Hadoop and Spark programming systems, as well as their supporting projects (e.g. HDFS, SparkSQL), have greatly simplified the analysis and transformation of datasets whose size exceeds the capacity of a single machine. While these programming models facilitate the use of distributed systems to analyze large datasets, they have been plagued by performance issues. The I/O performance bottlenecks of Hadoop are partially responsible for the creation of Spark. Performance bottlenecks in Spark due to the JVM object model, garbage collection, interpreted/managed execution, and other abstraction layers are responsible for the creation of additional optimization layers, such as Project Tungsten. Indeed, the Project Tungsten issue tracker states that the "majority of Spark workloads are not bottlenecked by I/O or network, but rather CPU and memory". In this work, we address the CPU and memory performance bottlenecks that exist in Apache Spark by accelerating user-written computational kernels using accelerators. We refer to our approach as Spark With Accelerated Tasks (SWAT). SWAT is an accelerated data analytics (ADA) framework that enables programmers to natively execute Spark applications on high performance hardware platforms with co-processors, while continuing to write their applications in a JVM-based language like Java or Scala. Runtime code generation creates OpenCL kernels from JVM bytecode, which are then executed on OpenCL accelerators. In our work we emphasize 1) full compatibility with a modern, existing, and accepted data analytics platform, 2) an asynchronous, event-driven, and resource-aware runtime, 3) multi-GPU memory management and caching, and 4) ease-of-use and programmability. Our performance evaluation demonstrates up to 3.24x overall application speedup relative to Spark across six machine learning benchmarks, with a detailed investigation of these performance improvements.