Data bus slicing for contention-free multicore real-time memory systems

Abstract

Memory access contention is one of the main contributors to tasks' execution time variability in real-Time multicores. Existing techniques to control memory contention based on time-sharing memory access do not scale well with increasing complexity of multicores, leading to a rapid increase of WCET estimates. This is due to fact that requests from different tasks interleave in the access to memory, and for each of its requests a task has to make worstcase time allowances to account for the memory state left by the previous request, that may belong to a different task. In this paper, we propose a memory organization that controls contention by dividing the data bus into narrower independent data buses, thus removing conflicts among different tasks accessing memory. While narrower data buses require extra transfers, they allow exploiting memory locality, hence only slightly affecting average performance. Our evaluation on a solid space case-study shows that the proposed memory organization provides contention-free memory access facilitating timing analysis and tightening WCET estimates.