An Agenda for Improved Evaluation of Supercomputer Performance (1986) / Chapter Skim
Currently Skimming:
3. The Current State of the Art in Evaluation of Supercomputers
3. The Current State of the Art in Evaluation of Supercomputers
Pages 21-30
The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.
From page 21... ...
These advances challenge existing methods of performance measurement because we cannot precisely forecast the nature of all the computational techniques that eventually will be used in these new applications. Nevertheless, performance measurement of existing systems and workloads provides valuable insight in improving current applications and guiding future developments.
|
From page 22... ...
Today it is axiomatic that supercomputer system performance is a function of problems, implementation, and algorithms. The current trend toward performance enhancement by using multiple processors (multiple vector processors for large applications)
|
From page 23... ...
, Los Alamos National Laboratory (in press) , and Moore and Bucher (1981~.
|
From page 24... ...
There is a clear need to start building major benchmark programs that can easily be transported, with appropriate algorithmic changes, to new systems. If performance measurement is to become a science, then the writing of benchmark programs should also be rationalized and not be dependent on ad hoc procedures to overcome the difficulty of FORTRAN for portability.
|
From page 25... ...
From this class of programs, it is also possible to determine the vector lengths at which half the peak performance is attained, the vector lengths at which the breakeven point for vector-scalar tradeoffs is reached, the asymptotic peak performance for vector operations, and the amount of work that must be performed between synchronization points in order to overcome the cost of communication in a multiprocessor system. However, the use of hardware demonstration programs for benchmarking and performance measurement is useful only at a very low level -- for machine designers and for providing parameters for use in higher level performance modeling.
|
From page 26... ...
Stripped-Down Versions of Major Application Packages Generally, stripped-down versions of major application packages will retain what we have labeled as basic routines, and they will permit some, but not all, of the interactions required between basic routines. They contain little of the pre- and post-processing required by full application packages and may or may not include interactions of CPU and I/O requirements.
|
From page 27... ...
Developmental Programs A final class, developmental programs, reflects the need to consider future workload directions and the need to admit novel approaches to algorithm and architecture design. As such, developmental programs would be used to try out new algorithms, to devise approaches appropriate for new architectures, and to experiment with them.
|
From page 28... ...
In spite of the wealth of work, not much has been directly applied to supercomputer systems, primarily because the tradeoffs in supercomputing generally involve maximizing computational performance instead of trying to optimize overall system performance. SUMMARY State-of-the-art practice in performance measurement generally involves a hierarchy of test programs wherein each level of the hierarchy has its unique strengths and weaknesses with respect to a specific set of applications or computational structures.
|
From page 29... ...
In press. Los Alamos National Laboratory Computer Benchmarking 1986.
|
From page 30... ...
Unclassified Release LA-UR-81-1977. Los Alamos, New Mexico: Los Alamos National Laboratory.
|
Key Terms
This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More
information on Chapter Skim is available.