Conclusions for sharing patterns analysis Sample Clauses

Conclusions for sharing patterns analysis. On average, about 75% of the blocks accesses and from 35% to 40% of blocks/pages (depending on the granularity) on the analyzed applications correspond to instruction blocks, which, unlike data blocks, do not require coherence maintenance. Moreover, despite the fact that it is detected a high sharing degree of data blocks among cores, indeed only SW blocks (just 40% of the total number of data blocks) require coherence. Considering accesses, SW blocks agglutinate 60%, on average, of the total number of accesses. That means that if snoopy-based coherence is used, the broadcast traffic reduction will be around 40%, while if directory-based coherence is used, then the benefits will increase. In order to ease data classification, different page granularity degrees can be used at the expense of causing a loss of accuracy due to block misclassification. We observed that in the case of block counting there is a high misclassification degree. As a counterpart, in the number of accesses the degradation of classifying block or pages is lower. Also, it is observed that SW blocks are not spread over all the pages, but they are indeed distributed between a limited number of pages. In particular, less than 40% of data pages require coherence. Deeping inside SW pages, we observe that they hardly are populated, containing about 30% and 10% of blocks, on average, for 4KB and 64KB pages, respectively. Among them, the number of SW blocks is the majority. Furthermore, about 30% of the accesses to blocks of SW pages correspond to store operations. Finally, it is observed that the impact of the producer-consumer sharing pattern in the analyzed applications is negligible. Less than 1,5% of the blocks accessed in SW pages present such a behavior, which discourages the application of update policies for coherence purposes in an extensive way. Further discussion must be done in order to decide which can be the best options among all the coherence protocols in the literature, taking into account also the characteristics of the target system of the study. As aforementioned, the impact of producer-consumer pattern is insignificant and therefore an invalidation-based protocol would be the best choice, which also will reduce the amount of cache-to-cache traffic and consequently reduce the amount of cache bandwidth and energy consumption. As seen on the analysis results, we found that for page granularities most blocks and accesses are classified as SW and, therefore, it may not be suita...