Next generation architectures and systems being deployed are characterized by high concurrency, low memory per-core, and multi-levels of hierarchy and heterogeneity. These characteristics bring out new challenges in energy efficiency, fault-tolerance and scalability. It is commonly believed that software has the biggest share of the responsibility to tackle these challenges. In other words, this responsibility is delegated to the next generation programming models and their associated middleware/runtimes. This workshop focuses on different aspects of programming models such as Task-based parallelism (X10, OCR, Habanero, Legion, Charm++, HPX), PGAS (OpenSHMEM, UPC, CAF, Chapel, etc.), Directive-based languages (OpenMP, OpenACC), Accelerator programming (CUDA, OpenCL), Hybrid MPI+X, etc. It also focuses on their associated middleware (unified runtimes, interoperability for hybrid programming, tight integration of MPI+X, support for accelerators) for next generation systems and architectures. Objective of ESPM2 workshop is to serve as a forum that brings together researchers from academia and industry to share knowledge and their experience, on working in the areas of programming models, runtime systems, compilation and languages, and application developers.
Topics of interest to the ESPM2 workshop include (but are not limited to):
Papers should present original research and should provide sufficient background material to make them accessible to the broader community.
Workshop Program | |
8:45 - 9:00 |
Opening Remarks |
9:00 - 10:00 |
Keynote Speaker: Katherine Yelick, UC Berkeley/LBNL Title: Solved Problems, Unsolved Problems, and Non-Problems in Exascale Programming Models Slides Abstract:
For several years hardware experts have argued that exascale systems will be
constrained by power density and total system energy, and will require new
programming models and implementation strategies to achieve good performance.
The programming models research community identified the need for fine-grained
parallelism, asynchronous execution, one-sided communication, autotuning,
communication-avoiding algorithms, dynamic scheduling, resilience and many other
exascale solutions and exascale systems will look like and what major
programming model challenges remain. The open problems include performance
portability, trading off locality and load balancing, and avoiding
synchronization and communication, along with correctness and resilience
concerns. But I will argue that the next phase in exascale computing research
needs to focus on science breakthroughs that will be enabled by these systems
and the programming model problems that arise from the science. The
applications are also evolving to use more complex multi-physics models and
multiscale methods, to incorporate observational data into simulations, and to
analyze complex scientific data sets.
|
10:00 - 10:30 |
Coffee Break |
10:30 - 12:30 |
Research Paper Session I : Emerging Programming Models Session Chair: Huo Zhigang - Institute of Computing Technology, Chinese Academy of Sciences, China PPL: An Abstract Runtime System for Hybrid Parallel Programming, Alex Brooks, Hoang-Vu Dang, Nikoli Dryden and Marc Snir - (University of Illinois at Urbana-Champaign) Slides ACPdl: Data-Structure and Global Memory Allocator Library over a Thin PGAS-Layer, Yuichiro Ajima, Takafumi Nose, Kazushige Saga, Naoyuki Shida and Shinji Sumimoto - (Fujitsu Limited) Slides Fault Tolerance Features of a New Multi-SPMD Programming/Execution Environment, Miwako Tsuji (RIKEN AICS), Serge Petiton (CNRS/LIFL) and Mitsuhisa Sato (RIKEN AICS/University of Tsukuba) Slides Higher-level Parallelization for Local and Distributed Asynchronous Task-Based Programming, Hartmut Kaiser (CCT/LSU), Thomas Heller, Daniel Bourgeois (CCT/LSU) and Dietmar Fey (Friedrich-Alexander-University) Slides |
12:30 - 2:00 |
Lunch Break |
2:00 - 2:30 |
Invited Talk Speaker: Olivier Tardieu, Research Staff Member T.J. Watson Research Center, Yorktown, NY, USA Title: Resilient and Elastic APGAS Slides Abstract: The APGAS programming model (Asynchronous Partitioned Global Address Space) is a simple but powerful model of concurrency and distribution, known primarily as the foundation of the X10 programming language but also developed for Java and Scala. APGAS combines PGAS with asynchrony. The data in an application is logically partitioned into places. The computation is organized into lightweight asynchronous tasks. APGAS can express both regular and irregular parallelism, within and across shared-memory nodes in a distributed system. Recently APGAS has been enriched to support failure-aware and elastic programming. Resilient applications can detect the loss of a place and implement recovery strategies. Elastic applications can dynamically add places to a running instance. In this talk, I will give an introduction to resilient and elastic APGAS, discussing design principles, implementation efforts, and applications. |
2:30 - 3:00 |
Research Paper Session II: GPU Computing Session Chair: Khaled Hamidouche - The Ohio State University The Scalable Petascale Data-Driven Approach for the Cholesky Factorization with Multiple GPUs, Yuki Tsujita (Tokyo Institute of Technology), Toshio Endo (Tokyo Institute of Technology/JST CREST) and Katsuki Fujisawa (Kyushu University/JST CREST) Slides |
3:00 - 3:30 |
Coffee Break |
3:30 - 4:00 |
Research Paper Session: Short Papers Session Chair: Takeshi Nanri - Kyushu University, Japan Hyper-Q-Aware Intranode MPI Collectives on the GPU, Iman Faraji and Ahmad Afsahi - (Queen's University) Slides Task Characterization-driven Scheduling of Multiple Applications in a Task-based Runtime, Kavitha Chandrasekar (Georgia Institute of Technology), Balasubramanian Seshasayee (Intel Federal), Ada Gavrilovska (Georgia Institute of Technology) and Karsten Schwan (Georgia Institute of Technology) Slides |
4:00 - 5:30 |
Panel Topic : Programming Models for Exascale Systems: What, When and How? Panel Moderator : Karl Schulz, Intel Corporation Panel Members:
|
5:30 - 5:40 |
Closing Remarks |
Coming Soon!