Adaptive Block-Wise Mapping With Intra-Block Resource Allocation for Multi-DNN Workloads on Heterogeneous Accelerator Systems

Deep neural networks (DNNs) dominate workloads on cloud and edge platforms. Meanwhile, the hardware platform towards the heterogeneous system with various accelerators. By mapping layers to their different preferred accelerators, the computation cost of each layer can be reduced. While mapping these layers on the same accelerator can reduce the inter-accelerator communication cost. These two costs are often competing and difficult to optimize simultaneously. Therefore, the core challenge in achieving efficient execution of DNN workloads on heterogeneous systems is: how to map layers to achieve the best trade-off between computation and communication costs. Existing works group layers into blocks and perform blockwise mapping to reduce inter-layer communication within blocks. However, when grouping layers, they typically rely on modelagnostic rules, which fail to hide critical inter-layer communication within blocks for diverse DNNs. Moreover, after block mapping, the lack of intra-block resource allocation further increases computation cost of block. In this paper, we proposeGHCoM, a novel block-wise mapping framework for exploring the effective cost trade-offs.GHCoMemploys an adaptive grouping strategy to guide layer grouping based on the topology of DNNs and dynamically adjust the grouping according to the trade-off target. Furthermore,GHCoMconsiders the fine-grained allocation of computation (i.e., processing elements) and communication (i.e., on-chip bandwidth) resources within each block to mitigate interlayer resource contention. To jointly optimize layer grouping, block-wise mapping and intra-block resource allocation,GHCoMleverages a two-level genetic algorithm (GA) with tailored encodings and operators that capture the interdependence across the entire design space. Experiments across various workloads and system configurations show thatGHCoMconsistently outperforms state-of-the-art baselines, achieving 1.08× to 4.79× speedup in execution latency and reducing energy consumption by 1.83% to 87.71%.