Designing GPU Architecture for Memory Bandwidth Reservation

Memory-Aware Latency Prediction Model for Concurrent Kernels in Partitionable GPUs: Simulations and Experiments

“…Aguilera et al [2] demonstrated the unfairness of spatial multitasking and proposed a fair resource allocation strategy for both performance and fairness. Emir C. Marangoz et al [15] showed that co-running applications can have interference in L2 cache, as we also highlighted in our work, and proposed a new architecture for managing BW reservation for a single kernel during concurrent execution. Unlike most of the research efforts we summarized in this section, our goal was to provide a predictive model for memory interference in cases of GPU SM spatial partitioning.…”

“…min(P enalties BW (k, n, w), I(w)) P enalties(k, n, w) < 1 P enalty M P D (k, n, w) P enalties(k, n, w) ≥ 2 P enalties(k, n, w) Otherwise (15) In order to leverage the 2x term in P enalties, we define a Max Penalty Density (P enalty M P D ), eq. (16a).…”