IAS Seminar: Towards an Execution Model-Driven IESP X-Stack Strategy

The International Exascale Software Project is the premier world-wide community forum for stimulating and coordinating cooperation among the European Union, the Pacific Rim, and the Americas towards the development of a shared body of software for systems of the next decade operating within the Exascale performance regime. The purpose of this pan-international body is to share ideas, resources, talents, and cooperative research programs (e.g., EESI) to realize a common framework for programming, compiling, runtime, and operating environments in preparation and time for the needs of Exascale systems by 2018 and beyond.

But what should that software look like, how should it perform, and how will it be used? More deeply, the controversial question is: what will be the future paradigm of parallel computing? A majority of the community favors either the continuance of current conventional practices or incremental extensions to them. This is a responsible view that reflects the history of the last two decades where such techniques have continued to serve. But, unfortunately, it fails to recognize the greater history of our field that has exhibited at least five epochs embodying distinct and distinguished paradigms, each with its own execution model. Each phase change to a new execution model has been driven by progress of underlying enabling technologies exhibiting new opportunities but also new challenges. Every execution model has provided a corresponding set of governing principles guiding the structure, management, and use of the hardware and software system so devised. What will that execution model be for the Exascale epoch?

ParalleX is an experimental execution model derived to support extreme scale computing both to reduce time to solution for strong scale problems and to fully employ resources for Exascale systems' performance. ParalleX integrates a number of research concepts in an innovative synthesis combined with new ideas to create a dynamic adaptive execution system with global address space and message-driven multi-threading operation. It employs powerful but lightweight synchronization semantics to eliminate global barriers and expose inherent parallelism derived from the meta data of complex structures.

This presentation will briefly describe ParalleX and present results from the HPX-3 experimental runtime system that shows performance benefits for applications employing dynamic graph data structures. ParalleX represents a class of execution models that may constructively serve as an alternative to conventional methods as the foundation for the X-stack to be developed under the auspices of IESP.

Anyone interested is cordially invited to participate in this seminar.