Final Report Summary - DEEPSEA (Parallelism and Beyond: Dynamic Parallel Computation for Efficiency and High Performance)
In this project, we developed techniques for efficient dynamic parallelism on modern multicore machine.
In one line of work, we developed a scheduling framework that takes advantage of the classic work-stealing technique but offers flexibility in implementing different policies for creating and
scheduling threads so that overheads of parallelism can be reduced without harming scalabality. We provided the theoretical analysis of this scheduling framework and provided several implementation for it in the C language. Based on this framework, we developed a range of techniques for supporting efficient parallelism with high-level language abstractions and applied these to several problems including in graphs. We developed language abstractions for parallelism, developed their semantics, and their cost model, and provided their implementation.
In another line of work. we developed techniques for enabling dynamic parallelism, where computations can respond to a wide variety of dynamic changes to their data automatically and efficiently, also in parallel. To this end, we extended the classic dynamic computation models to allow for bulk data changes and developed techniques for handling such changes efficiently in parallel. We developed language abstractions for writing such dynamic-parallel programs, specified their semantics, analyzed their efficiency, and provided several implementations. We also showed that these techniques could be integrated with existing industry-scale software such as the Apache Hadoop framework and can improve the efficiency of applications operating on large data sets.
In one line of work, we developed a scheduling framework that takes advantage of the classic work-stealing technique but offers flexibility in implementing different policies for creating and
scheduling threads so that overheads of parallelism can be reduced without harming scalabality. We provided the theoretical analysis of this scheduling framework and provided several implementation for it in the C language. Based on this framework, we developed a range of techniques for supporting efficient parallelism with high-level language abstractions and applied these to several problems including in graphs. We developed language abstractions for parallelism, developed their semantics, and their cost model, and provided their implementation.
In another line of work. we developed techniques for enabling dynamic parallelism, where computations can respond to a wide variety of dynamic changes to their data automatically and efficiently, also in parallel. To this end, we extended the classic dynamic computation models to allow for bulk data changes and developed techniques for handling such changes efficiently in parallel. We developed language abstractions for writing such dynamic-parallel programs, specified their semantics, analyzed their efficiency, and provided several implementations. We also showed that these techniques could be integrated with existing industry-scale software such as the Apache Hadoop framework and can improve the efficiency of applications operating on large data sets.