Traditionally in order to write applications that run across heterogeneous hardware, programmers must manually combine code for multiple programming models together in ad-hoc ways. Domain-specific languages (DSLs) offer an alternative approach, as high-level implicitly parallel domain abstractions can be transparently lowered to multiple heterogeneous architectures. Delite is a framework for building high performance DSLs embedded in Scala. This talk will describe how we have extended Delite to run single-source, implicitly parallel DSL applications across clusters of machines of CPUs and GPUs. Using examples from existing Delite DSLs for data querying and machine learning, we will show how we can achieve high performance with a flexible high-level programming model using a set of powerful compiler analyses and optimizations. Finally we will compare Delite's performance with Hadoop and Spark across of set of applications.

No outline is available