.png){kind=icon}
.png){kind=icon}
Return
Termination detection in data-driven parallel computations/applications
.png){kind=icon}
PRE
AI
DOI：10.1016/S0743-7315(03)00014-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
High-performance computing applications with data-driven communication and computation characteristics require synchronization routines in the form of eureka, barrier, or termination synchronization. In this paper, we consider termination synchronization for two different execution models, the AP and the APS model. In the AP model, processors are either active or passive and a passive processor can be made active by another active processor. In the APS model, processors can also be in a server state. A passive processor entering the server state does not become active again. In addition, a server processor cannot change the status of other processors. We describe and analyze solutions for both models and present experimental work highlighting the differences between the models. We show that in almost all situations the use of an AP algorithm to detect termination in an APS environment will result in loss of performance. Our experimental work on the Cray T3E provides insight into where and why this performance loss occurs. (C) 2003 Elsevier Science (USA). All rights reserved.
Keywords:
data-driven algorithms
termination detection
tree and ring networks
synchronization
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4
Papers:
3.8K
Citations:
4.8K
Organization
No organization information available