High Performance Virtual Machines -- 1996 DARPA ITO Summary

PROJECT SUMMARY

DARPA Order Number E313

Contractor:

PRINCIPAL INVESTIGATOR

     Andrew A. Chien
     Department of Computer Science
     University of Illinois
     1304 West Springfield Avenue
     Urbana, Illinois 61801
     Phone: 217-333-6844
     Fax:  217-244-6500
     Email: achien@cs.uiuc.edu

Co-PI's

     Daniel A. Reed and David A. Padua
     Department of Computer Science/University of Illinois
     Email: reed@cs.uiuc.edu
     Email: padua@cs.uiuc.edu

Related Information

Objective

Approach

High Performance Virtual Machines depend on building a uniform, portable abstraction -- a virtual machine -- with predictable, high performance characteristics. To successfully insulate application programs, a virtual machine must (1) deliver a large fraction of the underlying hardware performance, (2) virtualize resources to provide portability and to reduce the effort in building application programs, and (3) deliver predictable, high performance. The project is developing novel technology that leverages commodity components (hardware and software) to deliver high performance communication over cluster and wide area interconnects, predictable communication and computation, coordinated scheduling, and uniform access to resources (e.g. files, mass storage, embedded sensors). The HPVM project involves not only the development of novel communication, scheduling, and resource management technologies, but also dissemination of a series of software release which embody these ideas.

1996 Accomplishments

FY 1997 Plans

• Develop and distribute implementations of the Fast Messages 2.0 (low-level datagram and scatter/gather) and the Message Passing Interface (MPI) which run atop Myricom's Myrinet and in a Windows NT environment with high performance. This basic infrastructure will enable many HPC message passing applications to be cluster capable.
• Design, develop, and distribute global address space interfaces for clusters including a shmem put/get library (adapted from the Cray T3x) and Pacific Northwest Laboratory's Global Arrays. This implementation will support operation across Myricom's Myrinet and Windows NT. High performance implementation of these API's will increase the range of parallel applications that can efficiently execute on clusters.
• Document, robustify, and release HPVM 1.0; the first version and release of a High Performance Virtual Machine. This system will include the high performance communication interfaces above, appropriate glue software for making cluster use convenient (batch managers, job monitors, etc.), and execute with high performance within Windows NT and across Myricom's Myrinet.
• Explore the delivery of dynamic coscheduling technology in Windows NT and as appropriate build external modules and internal hooks to support efficient coscheduling of threads in a parallel computation (microsecond scale) while simultaneously supporting timesharing workloads across computing nodes in a cluster. Demonstrate a prototype which supports efficient coscheduling in Windows NT
• Obtain and explore implementation of High Performance Virtual Machine API's atop alternative cluster interconnects (Tandem's Servernet and Digital's Memory Channel). Evaluate interconnects and select those appropriate for HPVM implementation.
• Explore integration of HPVM access and communication implementations with emerging wide area metacomputing systems such as Globus and Legion. As appropriate, define interfaces which enable these systems to exploit HPVM's as high performance computing elements within the metacomputing infrastructure.

Technology Transition

Prepared 24 Oct 1996