Design and Implementation of Lazy Release Consistency Protocol on Cohesion

碩士 === 國立成功大學 === 電機工程研究所 === 83 === In this thesis, efforts are devoted to the design and implementation of lazy release consistency (LRC) on Cohesion which is a distributed shared memory system developed in our laboratory. Relea...

Full description

Bibliographic Details
Main Authors: Kean-Heng Lim, 林建興
Other Authors: Ce-Kuen Shieh
Format: Others
Language:en_US
Published: 1995
Online Access:http://ndltd.ncl.edu.tw/handle/39450376765615206039
Description
Summary:碩士 === 國立成功大學 === 電機工程研究所 === 83 === In this thesis, efforts are devoted to the design and implementation of lazy release consistency (LRC) on Cohesion which is a distributed shared memory system developed in our laboratory. Release consistency is a more relaxed protocol in comparison with the well-known sequential consistency and it can be implemented by either lazy or eager algorithm. The lazy release protocol guarantees the shared data becoming consistent at a processor when the processor acquires the synchronization object. It implies that no consistency is enforced for the shared accesses between the synchronization points. The lazy algorithm is better than the eager one because it requirs fewer messages sent across the interconnection network while maintaining the consistency. However, LRC needs more complicated data structures and operations to keep trace of the internal state of each page. Cohesion supports a simple thread model that makes the multithreaded parallel program development available. By exploiting the multithreading technique, we may overlap the computation and the network transfer to hide the latency incurred in sending data across network. However, the implementation of LRC in the multithreaded environment is more complicate and some of the policies must be reconsidered and be refined. We have implemented LRC on Cohesion without the assistant of special preprocessor or compiler. From experimental results, we find out that the false sharing is remedied, the number of message is reduced and a better performance can be achieved.