Cross-Partition Memory Interference in Software-Partitioned GPUs for Real-Time Tasks

Modern GPUs offer high computational throughput but are not designed to ensure spatial or temporal isolation between concurrent tasks. Only a few server-grade devices with multi-instance GPU technology support true hardware partitioning. As consumer and embedded GPUs now include many Streaming Multiprocessors (SMs), running multiple tasks concurrently has become common to improve utilization. Software-based SM-level spatial partitioning, where each partition is restricted to a fixed set of SMs, offers a simple and portable way to isolate concurrent tasks. Although in this method SMs are isolated, resources such as the L2 cache and DRAM remain shared, causing memory interference that can affect execution times and lead to timing unpredictability, often overlooked in real-time analysis. This paper empirically characterizes cross-partition interference on a COTS GPU under software-only SM-level partitioning. Using a mix of compute- and memory-intensive tasks, partitioned and non-partitioned configurations are compared by measuring both end-to-end response times and GPU execution times. The results indicate that SM-level partitioning largely confines interference for compute-bound tasks. However, memory-bound tasks continue to cause substantial interference across partitions, increasing the task execution time by as much as twofold. These findings indicate that memory interference is a critical challenge for achieving timing predictability in real-time GPU-based cyber-physical systems, underscoring the need to address memory interference to achieve predictable performance.