Tuning Parallel Applications with AIMS

Objectives: The objective of our efforts is to provide tools to help detect performance bottlenecks and suggest ways to eliminate them. Scalable visual representations of execution traces will also be developed, focusing specially on limiting the amount of information gathered and displayed and on supporting flexible trace-browsing capabilities. Besides porting AIMS to closely-coupled highly parallel testbeds (such as IBM's SP-2), we have also develop a version of AIMS that supports performance debugging under distributed/ heterogeneous (multi-programmed, multi-user) environments that exploit the portable message-passing interface and multitasking capabilities provided by PVM (Parallel Virtual Machine) 3.2.

Approach: The approach is to provide a set of tools responsible for: automatic instrumentation, monitoring, post-mortem analysis and visualization of program performance. These processes do not require the user to make changes to the application and will have easy-to-use interface involving minimal user interaction. We also provide a version of AIMS on workstation clusters running PVM.

It has long been understood that one of the most important decisions that a user makes in writing an efficient parallel programs is that of data distribution and alignment. We will address this issue in detail with the help of compiler support and demonstrate the feasibility and usefulness of data-oriented views for CFD applications developed by CAS researchers.

We also plan to implement a number of methods for limiting the volume of trace information gathered and displayed by AIMS. The results of our research will be used for developing fast and flexible trace-browsing capabilities for future AIMS releases.

Accomplishment: Based on version 2.2 of the Automated Instrumentation and Monitoring System (AIMS), Convex produced CXTRACE that runs on their HP cluster product as well as their supercomputer SPP-1. Two other computer companies (Intel and Meiko) are looking at commercialization possibilities with AIMS. AIMS has also been distributed to DOE laboratories, NASA centers, and a number of universities for teaching/ evaluation. In FY 94, enhancements made to AIMS include: (1) a new instrumentor based on the Sage/Sigma Tool-kit developed at Indiana University by Prof. Dennis Gannon, (2) prototype monitoring and display facilities for I/O, and (3) a version of AIMS was developed for PVM.

Significance: AIMS provides users with detailed information about the program execution, with little overhead and a simple user interface, to enable the tuning of their parallel applications on HPCC Testbeds. Status/Plan: We will continue to work closely alongside CAS application specialists in identifying tool features most useful for their work. We plan to port AIMS to the IBM SP-2 and continue extending its capabilities in conjunction with industrial and academic collaborators. Point of Contact:

Return to the Table of Contents

curator: Larry Picha