Monday: Paper Abstracts

Session 2 Track 1: Scalability and the road to exascale

Session 2 Track 2: Fault tolerance

Full paper 1: Space Performance Tradeoffs in Compressing MPI Group Data Structures
Sameer Kumar, Philip Heidelberger and Craig Stunkel

MPI libraries sometimes use O(N) data structures. The Blue Gene/Q machine has 16 GB memory per node, where only 512 MB is available per process at ppn=32. We present techniques to compress the dense O(N) mapping data structures that map the logical process ID to the global rank. Our techniques minimize topological communicator mapping state by replacing table lookup with a mapping function. We explore caching schemes to optimize overheads of the mapping functions for recent translations.

TOP_______________________________________________

Full paper 2: Modeling MPI Communication Performance on SMP Nodes: Is it Time to Retire the Ping Pong Test
William Gropp, Luke Olson and Philipp Samfass

The postal model of communication, T=alpha+beta n, is perhaps the most commonly used communication performance model, and is often used in developing and evaluating parallel algorithms. While it was an effective model when it was first proposed, with the advent of multicore nodes, the model has become a poor match to modern hardware. We show a simple 3-parameter model that better captures the behavior of current parallel computing systems, and demonstrate its accuracy on several systems.

TOP_______________________________________________

Full paper 3: Introducing Task-Containers as an Alternative to Runtime-Stacking
Jean-Baptiste Besnard, Julien Adam, Sameer Shende, Marc Pérache, Patrick Carribault, Julien Jaeger and Allen D. Malony

Many-core architectures pose new challenges to the MPI model which has been designed for distributed memory context, requiring a shift in its programming model with, for example, new paradigms such as Endpoints or Sessions. This paper extends C/C++ languages to define a task-containers through an open-source extended TLS library. Such facility aims at easing collaboration between MPI processes, programming models and collocated tasks in general, proposing an alternatives to runtime stacking.

TOP_______________________________________________

Full paper 4: The MIG Framework: Enabling Transparent Process Migration in Open MPI
Federico Reghenzani, Gianmario Pozzi, Giuseppe Massari, Simone Libutti and William Fornaciari

This paper introduces the mig framework: an Open MPI extension to transparently support the migration of application processes, over different nodes of a distributed High Performance Computing (HPC) system. The framework provides a mechanism on top of which suitable resource managers can implement policies to react to hardware faults, address performance variability, improve resource utilization, perform a fine-grained load balancing and power thermal management.

TOP_______________________________________________

Full paper 5: Architecting Malleable MPI Applications for Priority-driven Adaptive Scheduling
Pierre Lemarinier, Khalid Hasanov, Srikumar Venugopal and Kostas Katrinis

Future supercomputers will need to support both HPC applications and Big Data applications in a common environment. This motivates scheduling systems to support malleable jobs. It also calls for future HPC applications to provide some malleability. We present and compare two different approaches to support malleable MPI applications: 1) using checkpoint/restart, 2) using ULFM and its runtime. We examine their effects on application execution times as well as their impact on resource management.

TOP_______________________________________________

Full paper 6: Infrastructure and API Extensions for Elastic Execution of MPI Applications
Isaías Comprés, Ao Mo-Hellenbrand, Michael Gerndt and Hans-Joachim Bungartz

Dynamic Processes support was added to MPI in version 2.0 of the standard. This feature of MPI has not been widely used by application developers in part due to the performance cost and limitations of the spawn operation. In this paper, we propose an extension to MPI that consists of four new operations. These operations allow an application to be initialized in an elastic mode of execution and enter an adaptation window when necessary, where resources are incorporated into or released from the application’s world communicator. A prototype solution based on the MPICH library and the SLURM resource manager is presented and evaluated alongside an elastic scientific application that makes use of the new MPI extensions. The cost of these new operations is shown to be negligible due mainly to the latency hiding design, leaving the application’s time for data redistribution as the only significant performance cost.

TOP_______________________________________________

Edinburgh Images

Last updated: 27 Jul 2016 at 10:12