Metodología para la evaluación de prestaciones del sistema de entrada/salida en computadores de altas prestaciones

摘要

The increase in processing units, the advance in speed and compute power, and the increasing complexity of scientific applications that use high performance computing require more efficient Input/Output(I/O) Systems. The performance of many scientific applications is inherently limited by the input/output subsystem. Due to the historical "gap" between the computing and I/O performance, in many cases, the I/O system becomes the bottleneck of parallel systems. Identification of I/O factors that influence on performance can help us to hide this "gap". This situation lead us to ask ourselves the next questions: Must adapt Scientific applications to I/O configurations?, How do I/O factors influence on performance? How do you configure the I/O system to adapt to the scientific application? Answering these questions is not trivial. To use the I/O system efficiently it is necessary to know its performance capacity to determine if it fulfills I/O requirements of application. We present a methodology to performance evaluate of the I/O system on computer clusters. We propose the I/O system evaluation by taking into account requirements of the application and the I/O configuration. To do this we have defined a methodology with three stages: characterization, analysis of I/O and evaluation. Furthermore, we define a methodology to obtain the I/O model of parallel application that has been utilized to compare different I/O subsystems. It allows us to determine how much of the system's capacity is being used taking into account the I/O phases of application. The methodology can be used to select the configuration with less I/O time from different I/O configurations. The I/O model of application is defined by three characteristics: metadata, spatial global pattern and temporal global pattern. We have evaluated the I/O system utilization of different configurations by considering the I/O model of application.