On predictability and optimization of multiprogrammed caches for real-time applications

For almost two decades caches have been used in conventional computers for better system performance. The performance of cache memory is measured by its hit ratio, which is a probabilistic measure on the percent of memory accesses that can be intercepted by the cache memory to reduce the average memory access time. Despite its effectiveness, the impact of the “probabilistically” reduced task execution time by caches to the predictability of the real-time systems was not fully analyzed. This paper studies the effects of hit ratio on the timing predictability of real-time tasks based on trace driven simulation. In our extensive experiments on some selected SPEC benchmarks, the variation in task execution time due to the “probabilistic” cache hit ratio were studied and a statistical bound for the cached execution time is derived. We present cache partition optimization strategies for real-time multiprogramming applications where tasks are periodic and scheduled based on their priority. The cache is divided into variable sized partitions so that the total task utilization can be adjusted to be lower than a figure where a task set can be made schedulable using the minimum cache size. This is an improvement over the work proposed previously, in which only the total cache miss-ratio was minimized without considering the task deadlines. Two different partitioning strategies are studied: (a) only one task per partition, (b) tasks sharing partitions. An algorithm is presented to implement the partitioning strategies and its performance is evaluated through experimental study