Dataflow models for shared memory access latency analysis

Performance analysis of applications in multi-core platforms is challenging because of temporal interference while accessing shared resources. Especially, memory arbiters introduce a non-constant delay which signicantly in uences the execution time of a task. In this paper, we selected a prioritybased budget scheduler as memory arbiter which bounds temporal interference by construction and is well suited for bursty service provision. While existing performance analysis approaches assume a constant memory access latency leading to high overestimation, we propose in this paper a conservative data ow model for this scheduler, in which the history of memory accesses is considered. In a case study with an MP3-decoder for an ARM7 processor, we show that using a constant memory access latency for the selected scheduler results in an overestimation of three order of magnitudes. Compared to simulation, the proposed data ow model shows an overestimation of less than 3% while in previous work the overestimation was up to 104%. Furthermore, the proposed approach improves the performance by about 20% compared to a time-division-multiplex scheduler. Categories and Subject Descriptors: C.4 [Performance of Systems]: Computer Systems Organization - modeling techniques, performance attributes General Terms: Performance, Reliability, Verication. Keywords: real-time, multiprocessor performance analysis, memory accesses, data ow model.