Técnicas para Categorização de Padrões de Compartilhamento em Sistemas Software DSM
Maria Clicia Castro, Claudio Luis Amorim · 1999
ln this paper, we introduce a new categorization algorithm called RITMO to determime precisely sharing patterns in software distributed shared memory (DSM). RITMO is based on the page sharing state information generated at run-time by a finite state machine (FIESTA). RITMO categorizes dynamically pages according to their inherent sharing patterns in contrast with the induced sharing patterns used in current algorithms. The main advantage of this approach is to enable RITMO to categorize accurately irregular applications in which finegrain memory accesses often mask inherent sharing patterns making categorizations based on induced patterns inexact. Both FIESTA and RITMO are transparent to the programmer and do not require compiler or hardware assistance. Furthermore, they add no extra messages to the protocol and intrusion is kept extremely low. Through detailed execution-driven simulation of six benchmark applications running on TreadMarks with 8 processors we show that RITMO categorization is highly precise on classifying pages that are responsible for more than 93% of the application page faults, as either single-writer or migratory in tive out of six benchmarks. In both irregular applications tested, 50% (in Water-Nsq) and 8% (in Barnes2) of pages which correspond to 26% and 60% of the application page faults respcctively, have induced sharing patterns different from their inherent sharing patterns which were precisely determined by RITMO. Also, our results show that prefetching and forwarding could be efficiently implemented due to the RITMO accuracy, reducing remote data overheads from 14% (in BarnesTmk) up to 90% (in Em3d). As a result, the application execution time improved from 7% (Barnes2) to 38% (Em3d). Our conclusion is that RITMO based on FIESTA is a effcctive technique to improve software DSM performance in a large class of applications.