Next: Evaluation of HPC-Netlib
Up: Evaluation of High-Performance Computing
Previous: Approach
Our evaluation of PTLIB software covered both parallel debuggers and
performance analyzers. We give a detailed description of the evaluation
criteria below. Note that it is has been refined and expanded to a
level of detail to enable it to serve as an evaluation checklist.
- Performance
- Includes accuracy, efficiency, and scalability.
- Accuracy
-
A performance monitoring tool is accurate if it doesn't cause too
great a change in the behavior and timing of the program it is
monitoring.
- Efficiency
-
The software runs fast enough, in that slow speed does not make it
an ineffective tool.
- Scalability
-
A parallel tool is scalable if its overhead grows in a reasonable
manner with increases in system and problem sizes. In some cases,
linear growth may not be acceptable.
- Capabilities
-
The tool has adequate functionality to effectively accomplish its
intended tasks.
- Versatility
- Includes heterogeneity, interoperability, portability, and
extensibility
- Heterogeneity
-
A heterogeneous tool can simultaneously be invoked on and/or have
its components running on all platforms in a heterogeneous system.
- Interoperability
-
A parallel tool is interoperable if its design is based on open
interfaces and if it conforms to applicable standards.
- Portability
-
A parallel tool is portable if it works on different parallel
platforms and if platform dependencies have been isolated to
specific parts of the code.
- Extensibility
-
A performance analysis tool is extensible if new analysis methods
and views can be added easily.
- Maturity
- Includes robustness, level of support, and size of user base.
- Robustness
-
A parallel tool is robust if it handles error conditions without
crashing and by reporting them and recovering from them
appropriately.
- Level of support
-
- Size of user base
-
- Ease of use
-
The software has an understandable user interface and is easy to use
for a typical NHSE user.
Most of the software characteristics described in the criteria above
require qualitative rather than quantitative assessment. For this
reason reviewers assign the reviewed package a numerical score on each of
the characteristics.
Currently over 20 parallel debuggers and performance analyzers have been
evaluated according to the above criteria. These packages include
AIMS, DAQV, LCB, MQM, NTV, Pablo, Pangaea, Paradyn, ParaGraph,
ParaVision, PGPVM, PVaniM, TotalView, Upshot, VAMPIR, VT, Xmdb, XMPI,
and XPVM. We have solicited author feedback on these evaluations,
and the initial evaluations have been updated based on the feedback
received. Web access to the evaluations is available through the
NHSE homepage at http://www.nhse.org/
Next: Evaluation of HPC-Netlib
Up: Evaluation of High-Performance Computing
Previous: Approach
Jack Dongarra
Fri Nov 15 09:09:21 EST 1996