On the Influence of the Selection Function on the Performance of Fat-Trees

Gilabert, Francisco Jurado; Gómez, María E.; López, Pedro; Duato, José

doi:10.1007/11823285_91

Cited by 20 publications

(18 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Minimal adapting routing between a pair of nodes can be accomplished by sending the packet to one of the nearest common roots or nearest common ancestors (NCA) of source and destination, and from there to the destination. Each packet traverses through two phases, and ascending adaptive phase to get to one of the NCA, followed by a descending deterministic phase [22] [49].…”

Section: Routing In K-ary N-tree Networkmentioning

confidence: 99%

Predictive and Distributed Routing Balancing for High Speed Interconnection Networks

Castillo

Lugones

Franco

et al. 2011

2011 IEEE International Conference on Cluster Computing

View full text Add to dashboard Cite

EnglishIn high performance clusters, current parallel application communication needs, such as tra c pattern, communication volume, etc., change along time and are di cult to know in advance. Such needs often exceed or do not match available resources causing resource use imbalance, network congestion, throughput reduction and message latency increase, thus degrading the overall system performance. Studies on parallel applications show repetitive behavior that can be characterized by a set of representative phases. This work presents a Predictive and Distributed Routing Balancing (PR-DRB) technique, a new method developed to gradually control network congestion, based on paths expansion, tra c distribution, applications pattern repetitiveness and speculative adaptive routing, in order to maintain low latency values. PR-DRB monitors messages latencies on routers and saves the found solutions to congestion, to quickly respond in future similar situations. Tra c congestion experiments were conducted in order to evaluate the performance of the method, and improvements were observed. CastellanoEn los clusters de altas prestaciones, los requerimientos actuales de las comunicaciones de las aplicaciones, como el patr on de tr a co, el vol umen de comunicaciones entre otras, pueden cambiar a lo largo del tiempo y son dif ciles de predecir. Estas necesidades generalmente exceden o no se corresponden con los recursos disponibles realmente, lo cual conlleva a una situaci on de desbalanceo de los recursos, congesti on en la red, reducci on del throughput y un incremento considerable en los valores de latencia de los mensajes. Todo esto conlleva una degradaci on general del rendimiento de todo el sistema computacional. Los estudios de las aplicaciones paralelas demuestran que estas tienen un comportamiento repetitivo. Adem as, esta repetitividad puede detectarse y caracterizarse a trav es de unas fases representativas. Este trabajo propone un Algoritmo de Encaminamiento Predictivo y Distribuido (PR-DRB). Este nuevo m etodo propone controlar la congesti on de la red de manera gradual bas andose en la expansi on controlada de caminos, la distribuci on del tr a co, la repetitividad en las aplicaciones paralelas y el encaminamiento adaptativo especulativo; de manera a mantener los valores de latencia controlados. PR-DRB monitorea la latencia de los mensajes en los encaminadores y guarda las mejores soluciones adaptativas encontradas a una situaci on de congesti on. Esto se realiza de manera a re aplicar estas mejores soluciones de manera r apida ante situaciones similares futuras. Fueron desarrollados varios experimentos que generen congesti on de tr a co a n de evaluar el rendimiento de la propuesta, y se han logrado mejoras importantes en el rendimiento global del sistema.Palabras clave: Redes de interconexi on, Computaci on de Altas Prestaciones, Encaminamiento Predictivo, Encaminamiento Basado en las Aplicaciones.vi Català En els cl usters d'altes prestacions, els requeriments actuals de les comunicacions de les apl...

show abstract

Section: Routing In K-ary N-tree Networkmentioning

confidence: 99%

Predictive and Distributed Routing Balancing for High Speed Interconnection Networks

Castillo

Lugones

Franco

et al. 2011

2011 IEEE International Conference on Cluster Computing

View full text Add to dashboard Cite

show abstract

“…In [1], the authors compare several selection functions for adaptive routing, but they only consider the peak throughput in the evaluation, so they conclude that selection function is not critical. However, in [6], several selection functions for fat-trees were proposed and the authors notice that selection function has a high impact over average latency. For this, we only consider the two most representative selection functions proposed in [6].…”

Section: Network Modelmentioning

confidence: 99%

“…However, in [6], several selection functions for fat-trees were proposed and the authors notice that selection function has a high impact over average latency. For this, we only consider the two most representative selection functions proposed in [6]. The best one was Static And Destination Priority (SADP).…”

Section: Network Modelmentioning

confidence: 99%

Deterministic versus Adaptive Routing in Fat-Trees

Gómez

Gilabert

Gómez

et al. 2007

2007 IEEE International Parallel and Distributed Processing Symposium

Self Cite

123

View full text Add to dashboard Cite

show abstract

“…This selection function chooses the output port whose buffer has more free space, since it considers that this output port is less used than the other one, so the packet would be able to arrive sooner to its destination than using the other output port. This is only an example of selection function, as the selection functions may use very different criteria to make the final decision, as it can be seen in [50].…”

Section: Routing Algorithms Taxonomymentioning

confidence: 99%

“…In order to perform a fair comparison, we are not going to compare our deterministic routing algorithm against a single selection function. Indeed, we perform such comparison against several selection functions explained below, some of which are presented and compared in [50]. All these selection functions, but the last one, can be implemented as indicated above.…”

Section: Adaptive Routing Issuesmentioning

confidence: 99%

Low-Memory Techniques for Routing and Fault-Tolerance on the Fat-Tree Topology

Requena¹

View full text Add to dashboard Cite

On the Influence of the Selection Function on the Performance of Fat-Trees

Cited by 20 publications

References 8 publications

Predictive and Distributed Routing Balancing for High Speed Interconnection Networks

Predictive and Distributed Routing Balancing for High Speed Interconnection Networks

Deterministic versus Adaptive Routing in Fat-Trees

Low-Memory Techniques for Routing and Fault-Tolerance on the Fat-Tree Topology

Contact Info

Product

Resources

About