Publication: AdaptBenchmark : el benchmark adaptativo universal
Loading...
Identifiers
Publication date
2011-12
Defense date
2011-12-12
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Desde la primera aparición de las plataformas de computación, ha sido necesaria la comparación del rendimiento de los distintos productos. El término benchmark emerge a finales del siglo XIX, definido como un punto de referencia en la toma de medidas. Actualmente, este tema está más relacionado con herramientas software que evalúan el rendimiento del ordenador. Este proyecto trata sobre la definición, creación y evaluación de un conjunto de benchmarks para tomar medidas del rendimiento de los microprocesadores. Lo hemos denominado AdaptBenchmark: el benchmark adaptativo universal. Este benchmark incluye diferentes test que evalúan varias características de la arquitectura de un procesador. En este documento explicamos la metodología que hemos seguido durante el desarrollo del AdaptBenchmark. Detallamos la estructura interna del software, que ha sido concebida para evaluar el rendimiento de modernos procesadores con varios núcleos. AdaptBenchmark mide el tiempo de ejecución y eventos internos como: fallos caché en L1 y L2, tráfico de bus o fallos TLB. Esta información es recogida utilizando contadores hardware y resumida mediante gráficas. Para cada test desarrollamos un estudio extensivo del desarrollo de procesador así como un análisis de las interacciones entre software y el correspondiente hardware. Además, introducimos una técnica para buscar la mejor configuración en las distintas plataformas. Para ello utilizaremos dos técnicas de optimización: Simmulated Annealing y la Búsqueda N-aria. Esta última fue creada específicamente para este proyecto. Con la utilización de estas metodologías es posible adaptar el benchmark a distintos contextos hardware. A lo largo de este proyecto, trabajaremos con dos arquitecturas diferentes: Intel Core2Duo y AMD Opteron. Y compararemos al detalle el funcionamiento de estas dos plataformas para distintas configuraciones, probando así la eficiencia de nuestra herramienta. ________________________________________________________________________________________________________________
Since the first appearance of computing platforms, it was necessary to compare the performance of different products. The term benchmark emerges at the end of 19th century, defined as a point of reference for a measurement. Nowadays, this topic is commonly related to software tools that evaluate the computer performance. This project is about the definition, creation and evaluation of a set of benchmarks for measuring the microprocessor performance. We call them AdaptBenchmark: a universal adaptive benchmark. This benchmark includes different tests that evaluate several characteristics of the processor architecture. In this document we introduce the methodology that we have followed during the development of AdaptBenchmark. We detail the internal structure of this software, which was conceived for evaluating the performance of modern multicore processors. AdaptBenchmark measures the execution time as well as internal processor events like L1 and L2 cache misses, bus traffic or TLB misses. This information is collected using the hardware counters and then summarized in chars. For each test we perform an extensive study of the processor performance as well as an analysis of the interactions between the software and the underlining hardware. In addition, we introduce a technique for fine-tuning the benchmark in order to find the best configuration for each platform. In order to do this, we use two optimization techniques: the Simmulated Annealing and the n-ary Search. The latter one was originally created in this project. With the use of this methodology it is possible to adapt the benchmark to different hardware contexts. Finally, we perform the evaluation of AdaptBenchmark for two different architectures: Intel Core2Duo and AMD Opteron. We make a detailed comparison between these two platforms for different hardware configurations, proving the efficiency of our tool.
Since the first appearance of computing platforms, it was necessary to compare the performance of different products. The term benchmark emerges at the end of 19th century, defined as a point of reference for a measurement. Nowadays, this topic is commonly related to software tools that evaluate the computer performance. This project is about the definition, creation and evaluation of a set of benchmarks for measuring the microprocessor performance. We call them AdaptBenchmark: a universal adaptive benchmark. This benchmark includes different tests that evaluate several characteristics of the processor architecture. In this document we introduce the methodology that we have followed during the development of AdaptBenchmark. We detail the internal structure of this software, which was conceived for evaluating the performance of modern multicore processors. AdaptBenchmark measures the execution time as well as internal processor events like L1 and L2 cache misses, bus traffic or TLB misses. This information is collected using the hardware counters and then summarized in chars. For each test we perform an extensive study of the processor performance as well as an analysis of the interactions between the software and the underlining hardware. In addition, we introduce a technique for fine-tuning the benchmark in order to find the best configuration for each platform. In order to do this, we use two optimization techniques: the Simmulated Annealing and the n-ary Search. The latter one was originally created in this project. With the use of this methodology it is possible to adapt the benchmark to different hardware contexts. Finally, we perform the evaluation of AdaptBenchmark for two different architectures: Intel Core2Duo and AMD Opteron. We make a detailed comparison between these two platforms for different hardware configurations, proving the efficiency of our tool.
Description
Keywords
Benchmark, Herramientas de software, Microprocesadores, Control de calidad