Monitoring and Debugging Parallel Software with BCS-MPI on Large-Scale Clusters

Authors:
Juan Fernandez;Fabrizio Petrini;Eitan Frachtenberg
Affiliations:
Universidad de Murcia, Spain;Los Alamos National Laboratory, NM;Los Alamos National Laboratory, NM
Venue:
IPDPS '05 Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS'05) - Workshop 18 - Volume 19
Year:
2005

Citing 9
Cited 0

Dynamic software testing of MPI applications with umpire

Proceedings of the 2000 ACM/IEEE conference on Supercomputing
Predictive performance and scalability modeling of a large-scale application

Proceedings of the 2001 ACM/IEEE conference on Supercomputing
The Quadrics Network: High-Performance Clustering Technology

IEEE Micro
STORM: lightning-fast resource management

Proceedings of the 2002 ACM/IEEE conference on Supercomputing
Replay for Debugging MPI Parallel Programs

MPIDC '96 Proceedings of the Second MPI Developers Conference
Architectural Support for System Software on Large-Scale Clusters

ICPP '04 Proceedings of the 2004 International Conference on Parallel Processing
The Case of the Missing Supercomputer Performance: Achieving Optimal Performance on the 8,192 Processors of ASCI Q

Proceedings of the 2003 ACM/IEEE conference on Supercomputing
BCS-MPI: A New Approach in the System Software Design for Large-Scale Parallel Computers

Proceedings of the 2003 ACM/IEEE conference on Supercomputing
Scalable NIC-based Reduction on Large-scale Clusters

Proceedings of the 2003 ACM/IEEE conference on Supercomputing

Quantified Score

Hi-index	0.00

Visualization

Abstract

Buffered CoScheduled (BCS) MPI is a novel implementation of MPI based on global synchronization of all system activities. BCS-MPI imposes a model where all processes and their communication are tightly scheduled at a very fine granularity. Thus, BCS-MPI provides a system that is much more controllable and deterministic. BCS-MPI leverages this regular behavior to provide a simple yet powerful monitoring and debugging subsystem that streamlines the analysis of parallel software. This subsystem, called Monitoring and Debugging System (MDS), provides exhaustive process and communication scheduling statistics. This paper covers in detail the design and implementation of the MDS subsystem, and demonstrates how the MDS can be used to monitor and debug not only parallel MPI applications but also the BCS-MPI runtime system itself. Additionally, we show that this functionality need not come at a significant performance loss.