A Study Of The Pre-evaluation Methods On Performance Balance Of Supercomputer

The peak computing capability of supercomputer systems keep on enhancing while the practical applications of them are still inefficient. With the growing contradictions between them, research on the performance balance of supercomputer systems is causing widespread concern in the industry gradually. It is helpful to prejudge the suitable workload type of the system and improve the efficiency of the system's practical applications by researching the pre-evaluation methods on performance balance of supercomputer system. The research can also dig out the performance bottleneck and structure flaw in the process of architecture designing, then lead the shorting of cost and risk.In this essay, the current evaluating methods of the supercomputer systems' performance balance, which is based on experience and benchmark routine are analyzed, and their deficiencies on pre-evaluation and integration evaluation are pointed out. The essay raises the performance balance pre-evaluation method which is based on virtual stimulating and simulation-based prediction technology. By describing the features of workload and analyzing the typical application workload, an instruction-level workload model is constructed and the virtual stimulate of typical application workload is generated. The concept calls unbalance-degree is put forward referencing the theories in multi-target complex evaluating. This provides a quantitative method on performance balance evaluating. Also, a performance balance pre-evaluating system is set up. In this system, the simulator of supercomputer performance is used as a performance prediction platform, virtual stimulate is used as the simulator driver, the evaluating standard is the balance requirement, and unbalance-degree is used as the quantitative method.Finally, the Blue Gene/L is used as the target machine. Its performance balance is evaluated by using the performance balance pre-evaluating system. Through the evaluating, the system's suitable workload type is determined, and the bottleneck of the system is pointed out. Improvements are put forward and the effectiveness of the balance performance pre-evaluating system is proved.