Data Driven Network of Workstations (D2NOW)

Abstract

This paper presents the Data Driven Network Of Workstations (D2NOW), a multithreaded architecture that is based on the Decoupled Data Driven model of execution. This model decouples the synchronization from the computation portions of a program and allows them to execute asynchronously. At compile time a Multithreaded program is created with a Data-Driven thread synchronization graph superimposed on it. D2NOW is built using commodity control-flow microprocessors. The support for the data driven synchronization of threads, is provided by the Thread Synchronization Unit (TSU). The TSU is attached in the COAST (Cache On A STick) L2 Cache slot of Pentium workstations and thus it has an implicit interface, using snooping, to the Pentium microprocessor. Workstations are connected via a Telegraphos interconnection network, which is a high throughput ATM-like switch. Telegraphes uses short packets and guarantees no packet-drop, which is a must for fine grain data-driven computation. D2NOW exhibits the tolerance to long memory and communication latencies, of the data-driven model, with very little overhead and also exploits short-term optimal cache placement and replacement policies. In our prototype implementation the TSU is implemented using FPGAs and it has very low hardware overhead. © Springer Pub. Co.