19

Fig. 3.1. The PRESS Model

receiving messages. These threads remain blocked until there is a message to be sent

or received. When the receive thread wakes up, it determines which virtual interface

(VI) [30] receives the message and places a descriptor of the message in the queue shared

with the main thread. The main thread periodically polls this data structure [23]. To

forward requests based on the cache locality and load balance, this information should be

shared by all the nodes in the cluster. In PRESS, the cache information is broadcasted

whenever a cache is replaced. It employs broadcasting and piggy backing alternatively for

disseminating the load balance information. The PRESS design has been implemented

on an 8 node Linux cluster connected with a Giganet (cLAN) network that used VI as

the communication substrate. Measurement with real workloads provided about 29%

improvement in throughput compared to the TCP/IP implementation.