Empirical Bounds on Data Caching in High- Performance Real-Time Systems

can be an order of magnitude longer than the best-case execution time; in the worst-case an instruction can result in two cache misses which can easily stall the processor for tens of cycles. At the same time, cache memories often handle a vast majority of all memory accesses in a single cycle. This has motivated us to develop methods to predict data cache behavior so as to provide tighter bounds on the worst-case execution time. The challenge is to determine the set of data accesses that are independent of input data. Such memory instructions are allowed to cache data. Recently a few studies have been published on methods to allow for predictable data caching. The basic method in [7] handles all unpredictable data accesses as if they result in two cache misses because in the worst case, a memory instruction will miss and replace data. If the predictable hit rate is lower than 50%, this method will produce more pessimistic results than if data caches were not used. Our approach is instead to let all memory accesses that are not predictable bypass the cache. Support for this exists in most embedded microprocessors. [6] is concerned with the problem of how to model conflicts between predictable data cache accesses and does not address the problem of how to handle unpredictable data accesses. Ultimately, the effectiveness of data caches in hard real-time systems is dictated by the fraction of all memory accesses that can be predicted, i.e., are input-data independent. The purpose of this paper is to make an estimation of this fraction by (1) formulating what predictable data caching is and by (2) providing empirical data on what fraction of data accesses that are predictable. We do this by analyzing a set of non-trivial programs from the SPEC95 benchmark suite. While our study is preliminary, our empirical data so far look promising; more than 84% of the data accesses are indeed predictable. This suggests that data caching is effective in hard real-time systems although the challenge is to find tractable methods to come close to this bound. The fraction of such predictable data accesses that can be covered by such methods establishes a practical bound on the effectiveness. In Section 2, we provide our approach to predictable data caching. Sections 3 and 4 present the experimental results and, finally, Section 5 discusses our ongoing work. Abstract In this paper we study …