Homepage of Technische Universität Dresden
Personal tools
|
|
Analysis of Architecture and Performance of High Performance ComputersWhile high-performance computers are becoming increasingly powerful, especially in terms of theoretical peak performance, achieving the best performance of scientific applications requires a thorough understanding of the architecture and performance characteristics of the computing system. To gain a useful understanding of system performance, which is an increasingly important aspect new system procurements, the use of accurate benchmarking tools is required.
The ZIH is engaged in several benchmarking projects and in the investigation of new computer architectures;
The BenchIT Project was started in 2002. Its pursuit is using an expandable set of interactive measurement programs to conclusively test a system's performance in a relatively short amount of time. Time-based and performance counter measurements can be used to assess performance. It is also possible to correlate multiple results during a performance run by applying one variable parameter. One of the design principles of BenchIT is a high degree of portability for a number of UNIX derivatives. [more]
The PARbench benchmark system was developed to examine how a multiprogramming environment influences the manner in which jobs are executed on a high performance computer. The first version was created 1988-1990, then further improved in 2001 and again in 2005 to match the conditions of modern supercomputers. Both serial and parallel jobs can be generated with given parameters by PARbench. Essential parameters for the serial case are the number of floating point operations per second (FLOPS), the number of memory references per second (MREFS) and the number of I/O operations per second (IOOPS). In addition to this, the degree of parallelism, the degree of load balancing and the number of usable CPUs are important for parallel workloads. The generated workloads are executed by the operating system as normal jobs. PARbench controls the logging of necessary data on its own. The analysis of the data files created by the benchmark run is managed by the PARbench visualization tool.
The ZIH is investigating the performance of new computing systems using cell processors. [more]completed projects:
|
NewsContactVisitor's address:
|
||||||||||||||||||||||||||||||||||||
beratung@zih.tu-dresden.de