PS-Architecture: A scalable and energy-efficient architecture for CMP NUCAs
As the number of cores increases in both incoming and future shared-memory chip--multiprocessor (CMP) generations, coherence protocols and all elements in the cache hierarchy must scale to sustain performance. In this work we attack the scalability problem in the CMPs by studying and proposing some...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2013 |
| País: | España |
| Institución: | Universitat Politècnica de València (UPV) |
| Repositorio: | RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia |
| Idioma: | inglés |
| OAI Identifier: | oai:riunet.upv.es:10251/44383 |
| Acceso en línea: | https://riunet.upv.es/handle/10251/44383 |
| Access Level: | acceso abierto |
| Palabra clave: | Cache Directory CMP ARQUITECTURA Y TECNOLOGIA DE COMPUTADORES Máster Universitario en Ingeniería de Computadores-Màster Universitari en Enginyeria de Computadors |
| Sumario: | As the number of cores increases in both incoming and future shared-memory chip--multiprocessor (CMP) generations, coherence protocols and all elements in the cache hierarchy must scale to sustain performance. In this work we attack the scalability problem in the CMPs by studying and proposing some improvements for two of those elements, namely the directory and data caches. Each of these two structures have its particular issues which we try to solve employing some mechanisms involving the different type of blocks that can be found in parallel workloads. We introduce the PS directory, a directory cache that uses two different cache structures, each one tailored to one of these types of blocks (i.e., private and shared). The Shared directory cache, which tracks shared blocks is small, with low associativity and fast. The Private directory cache is aimed at tracking private blocks, which are highly dominant in current workloads. This structure does not store the sharer vector, is larger than the shared cache, and it has higher associativity. We also introduce the PS cache, an energy-efficient cache design which only accesses a subset of the set ways without hurting performance. |
|---|