Compiler and memory layout in Fresh Breeze system

Date
2016
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
In the past years, processors have evolved into multi-core, and programs also turn into multi-threading. To meet the demand of parallel processing, we face new challenges to explore the implicit parallelism in programs and ease the programming on large scale parallel system. The MIT Fresh Breeze system applies principles of data flow computing, leverages the write-once and chunk based memory model to eliminate the cache coherence issue in modern multi-core processor. We develop a template based transform component in the Fresh Breeze compiler. It enables compiling funJava (subset of Java language) programs into codelets, automatically adapts the linear space memory model to the Fresh Breeze memory model. We test three common algebra function, demonstrate the Fresh Breeze system's massive concurrent execution ability and achieve the sub linear speedup performance. In addition, through evaluations by tuning the major features, we find optimal configurations for fresh breeze processor design. Comparison on various combinations of memory layouts proves that the conventional layout optimization techniques are able to effectively apply on the Fresh Breeze system. It brings future work on the global layout optimization for a group of functions in applications.
Description
Keywords
Citation