A Method to Reduce the Acknowledgement Overhead of S-DSM Systems
Kenji Kise, Takahiro Katagiri, Hiroki Honda, Toshitsugu Yuba · 2005
We discuss the inter-process communication in software distributed shared memory (S-DSM) systems. Some S-DSM systems, such as TreadMarks and JIAJIA, adopt the user datagram protocol (UDP) which does not provide the reliable communication between the computation nodes. To detect and recover from a communication error, therefore, an acknowledgment (ACK) is used for every message transmission in the middleware layer. In this paper, firstly, we show that an acknowledgement is not necessarily required per one message transmission in the middleware layer. Secondly, the method to reduce the acknowledgement overhead for a page request is proposed. We implemented the proposed method in our S-DSM system Mocha. The performance of the method was measured with several benchmark programs. We show that Matrix Multiply (MM) of high page transfer frequency achieves a drastic speedup as much as 92% for a 16-node PC cluster, compared with the conventional communication method of not omitting the acknowledgment.