A Field Practical Approach to Memory Bandwidth Allocation for Consolidating Multi-Domain Automotive Applications on a Single SoC

Along with the advent of a high-end SoC with multiple CPU clusters and many GPUs, the automotive industry has a strong motivation to consolidate multi-domain applications on such a single SoC for wiring harness reduction and space/weight saving. For this, it is essential to bound their mutual interferences among CPU+GPU clusters on the system memory since advanced automotive applications use huge size code/data like autonomous driving and multi-screen infotainment. In order to guarantee memory bandwidth to each cluster, this paper proposes a cluster-level memory access regulation that aggregates the total memory access amount by each cluster (by all CPU cores and GPUs within the cluster) and throttles the cluster if it exceeds the given threshold. Then, we propose a few-shot measurement based optimal memory bandwidth allocation that can find a near optimal solution with only a few measurements, which is practically essential to save the system development cost. Our extensive experiments on a real SoC board say that our proposed techniques successfully regulate each cluster's memory bandwidth usage within ± 10 % margin of the allocated bandwidth. Also, our optimization can achieve a near-optimal utility with less than 0.6 % loss on average compared to the real optimal, with only a few measurements (mostly three measurements) instead of 25 measurements needed for the real optimal.