Static cache partitioning robustness analysis for embedded on-chip multi-processors

In this paper we analyze the robustness of multi-tasking applications when mapped on an on-chip multiprocessor platform. We assume a multiprocesso structure which embeds a hierarchical cache organization with two levels. The first one is private to the processor cores while the second one is shared among the processors. To enable compositionality, i.e, to be able to evaluate the system's performance out of the individual task's performance, the second level of cache (L2) is partitioned per task basis. Two robustness aspects are relevant in this context: internal (performance deviations are caused by the tasks comprising the application) and external (performance variations are caused by external stimuli). First we introduce two metrics to quantify the robustness. The internal robustness is estimated by a sensitivity function which measures the performance variations induced by the %private inter-task cache interference. The external robustness is quantified by a stability function which reflects the variations induced by different input data on the partitioned L2 behavior. Subsequently, we exercise our method on two applications (H.264 and picture-in-picture TV) running on a CAKE multi-processor platform. Our experiments indicate that, if the cache is partitioned, the sensitivity is 8% and 5% for the H.264 and PiPTV, respectively. For the shared cache scenario the sensitivity is 40% and 50% for the H.264 and PiPTV, respectively. The variations induced in the L2 behavior by various input data sets are at most 4% for the PiPTV application, respectively 9% for the H.264 decoder. This accounts for a stability of at least 96%, respectively 91%, therefore, for the investigated applications, we can conclude that the static cache partitioning is quite robust to input stimuli.