Introducing runahead threads
Simultaneous Multithreading processors share their resources among multiple threads in order to improve performance. However, a resource control policy is needed to avoid resource conflicts and prevent some threads from monopolizing them. On the contrary, resource conflicts would cause other threads...
| Autores: | , , , |
|---|---|
| Tipo de documento: | relatório científico |
| Data de publicação: | 2007 |
| País: | España |
| Recursos: | Universitat Politècnica de Catalunya (UPC) |
| Repositório: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglês |
| OAI Identifier: | oai:upcommons.upc.edu:2117/108084 |
| Acesso em linha: | https://hdl.handle.net/2117/108084 |
| Access Level: | Acceso aberto |
| Palavra-chave: | Simultaneous multithreading processors Parallel processing (Electronic computers) Multiprocessadors Processament en paral·lel (Ordinadors) Àrees temàtiques de la UPC::Informàtica::Arquitectura de computadors |
| Resumo: | Simultaneous Multithreading processors share their resources among multiple threads in order to improve performance. However, a resource control policy is needed to avoid resource conflicts and prevent some threads from monopolizing them. On the contrary, resource conflicts would cause other threads to suffer from resource starvation degrading the overall performance. This situation is especially sensitive for memory bounded threads, because they hold an important amount of resources while long latency accesses are being served. Several fetch policies and resource control techniques have been proposed to overcome these problems by limiting the per-thread resource utilization. Nevertheless, this limitation is harmful for memory bounded threads because it restricts the memory level parallelism available that hides the long latency memory accesses. In this paper, we propose Runahead threads on SMT scenarios as a valuable solution for both exploiting the memory-level parallelism and reducing the resource contention. This approach switches a memory-bounded eager resource thread to a speculative light thread, avoiding critical resource blocking among multiple threads. Furthermore, it improves the thread-level parallelism by removing long-latency memory operations from the instruction window, releasing busy resources. We compare an SMT architecture using Runahead threads (SMTRA) to both state-of-the-art static fetch and dynamic resource control policies. Our results show that the SMTRA combination performs better, in terms of throughput and fairness, than any of the other policies. |
|---|