Unified Trace Environment
UTE is a unified trace environment designed to generate, analyze,
and visualize trace events for applications running on IBM
Scalable Parallel systems. No source code modification is
necessary for trace generation. Message passing events can be
generated along with system responses, such as process dispatch,
page fault, and I/O activities. User markers are provided to mark
various phases, loops, and routines if the application source code
is available. All these events can be visualized, not only for users to
understand the communication patterns of an application,
but also to understand system responses and pinpoint the
bottleneck of the application through source code visualization.
UTE is built on top of the IBM AIX trace facility for generating system
events. This provides a unified and easily expandable trace environment
for various software layers such as MPL, MPI, PIOFS, and HPF.
In IBM SP systems with POE 1.2 and AIX 3.2.5,
two trace libraries, UTE/MPIF and UTE/MPL, have been developed for MPIF and MPL
applications to generate message passing and system events
with minimum overhead (a few microseconds
per message passing event). Additional hooks have been inserted in other
software layers such as PIOFS and Tiger Shark for trace generation, making
UTE an easily expandable trace environment for IBM Scalable Parallel systems.
In IBM SP systems with POE 2.1 and AIX 4.1, a unified trace library is also
being developed for both MPI and MPL applications.
``A Distributed Connection Manager Interface for Web
Services on IBM SP Systems,'' Proc. of the 1996
International Conference on Parallel and Distributed Systems, Tokyo, June 1996.
This paper describes a distributed connection manager daemon for Web services
and a customized UTE trace library for such a client/server application
``A Scalable Web Server and Its Performance Analysis on Multiple Platforms,''
Proc. of ICDCS'96, Hong Kong, May 1996.
``UTE: A Unified Trace Environment for IBM SP Systems,''
Proc. of the 8th Conference on Parallel and Distributed
Computing Systems, pp. 540 - 547, September 1995. This
UTE paper describes the UTE/MPI and
UTE/MPL trace libraries along with tools
``MPI Programming Environment for IBM SP1/SP2,''
Proc. of ICDCS'95, May 1995. This paper
describes the MPI-F programming environment with UTE features
``Timestamp Consistency and Trace-Driven Analysis for
Distributed Parallel Systems,'' Proc. of IPPS'95, pp. 680 - 686, April 1995.
This paper describes the UTE/MPLp
trace environment and related issues on timestamp concistency
``Parallel I/O Workload Characteristics Using Vesta,''
Proc. of the IPPS'95 Workshop on Input/Output in Parallel and Distributed
Systems, pp. 16 - 29, April 1995, also IBM Research Report RC 19940, February
``Trace-Based Analysis and Tuning for Distributed Parallel Applications,''
Proc. of International Conference on Parallel and Distributed Systems,
pp. 716 - 723, December 1994. This paper describes the UTE/MPLp
trace environment and issues on performance tuning
C. Eric Wu
IBM home page |
Contact IBM |