References And911 P. Andersson. Performance Evaluation of Diierent T Opologies for the Data Diiusion Machine. Final Work for Undergraduate Studies, Kth, November

resulted in a poor speedup for the DDM as well as for other architectures. This poor locality w as tamed by restructuring its distribution of work from being static to being dynamic. The restructuring also increased MP3D's communication locality. C o m m unication locality w as also enhanced for Cholesky by adding hierarchical knowledge to its dynamical scheduler. Finally, the importance of the attraction memory was shown to increase with a larger data set. The conclusion is that a COMA can successfully execute shared-memory applications written with a diierent a r c hitecture in mind. Further improvements can be obtained by small modiication to the applications to better suit the unique properties of a COMA. 8 Acknowledgements Part of the DDM work has been performed under the ESPRIT project 2741 PEPMA. We thank the many colleagues involved in or associated with the project. cache-miss rates as the problem set is increased for the studied applications. Node Miss Hit Attr. Mem. Hit 1st Level Fig. 8. The eeects of the large attraction memories increase with the size of the data set. Here, the behavior of a small data set to the left is compared to larger data sets to the right for some applications. For the MP3D applications, the number of particles is increased while the number of space cells is held constant. 6 Related Work The hierarchical DDM and its protocol have s e v eral similarities with the archi-tectures proposed by Wilson Wil86, and Goodman GW88. The problem of maintaining coherence in hierarchical systems is addressed in the Wilson. The Goodman proposal has a network built by a grid of buses. A node snoops two buses and has two assignments in that architecture. First it is a processing node and secondly it acts as a directory. The DDM is diierent in its use of transient states in the protocol, its lack o f p h ysically shared memory, and its storage of state information but no data in the network higher level caches. It is also diierent in that it handles replacement i n s u c h a w ay that at least one copy o f all the data in the caches is always guaranteed, and thus the shared memory is not needed. Recently, the commercial machine KSR1 was released BFKR922. It is similar to the DDM being a hierarchical COMA, but has a much larger …