Deep Neural Networks (DNNs) have become increasingly capable of performing tasks ranging from image recognition to content generation.The training and inference of DNNs heavily rely on GPUs, as GPUs' massively parallel architecture delivers extremely high computing capability.With the growing complexity of DNNs and the size of training datasets, training DNNs with a large number of GPUs is becoming a prevalent strategy.Researchers have been exploring how to design software and hardware systems for GPU farms to achieve the best utilization, efficiency, and DNN accuracy during training or inference.However, when designing and deploying such systems, designers usually rely on testing on physical hardware platforms equipped with many GPUs, incurring high costs that are almost prohibitive for system designers to test different configurations and designs, even for highly resourceful companies.While an alternative solution is to test on GPU simulators, they are often too slow for these large-scale systems and depend on profiling details collected from real distributed systems to initiate the simulation.To address these challenges, we present TrioSim, a novel lightweight simulator for DNNs on multi-GPU systems.TrioSim combines performance modeling techniques and simulation methods to achieve high flexibility, high simulation speed, and * Part of this work was done while Yuhui Bao and Pranav Vaid were interns at Lightmatter.

Energy conservation of large data centers for high performance computing workloads, such as deep learning with Big Data, is of critical significance, where cutting down a few percent of electricity translates into million-dollar savings. This work studies energy conservation on emerging CPU-GPU hybrid clusters through dynamic voltage and frequency scaling (DVFS). We aim at minimizing the total energy consumption of processing a batch of offline tasks or a sequence of real-time tasks under deadline constraints. We derive a fast and accurate analytical model to compute the appropriate voltage/frequency setting for each task, and assign multiple tasks to the cluster with heuristic scheduling algorithms. In particular, our model stresses the nonlinear relationship between task execution time and processor speed for GPU-accelerated applications, for more accurately capturing real-world GPU energy consumption. In performance evaluation driven by real-world power measurement traces, our scheduling algorithm shows comparable energy savings to the theoretical upper bound. With a GPU scaling interval where analytically at most 36% of energy can be saved, we record 33-35% of energy savings. Our results are applicable to energy management on modern heterogeneous clusters.

Conserving the energy consumption of large data centers is of critical significance, where a few percent in consumption reduction translates into millions-dollar savings. This work studies energy conservation on emerging CPU-GPU hybrid clusters through dynamic voltage and frequency scaling (DVFS). We aim at minimizing the total energy consumption of processing a sequence of real-time tasks under deadline constraints. We compute the appropriate voltage/frequency setting for each task through mathematical optimization, and assign multiple tasks to the cluster with heuristic scheduling algorithms. In performance evaluation driven by real-world power measurement traces, our scheduling algorithm shows comparable energy savings to the theoretical upper bound. With a GPU scaling interval where analytically at most 38% of energy can be saved, we record 30-36% of energy savings. Our results are applicable to energy management on modern heterogeneous clusters. In particular, our model stresses the nonlinear relationship between task execution time and processor speed for GPU-accelerated applications, for more accurately capturing real-world GPU energy consumption.

Memory access efficiency is a key factor in fully utilizing the computational power of graphics processing units (GPUs). However, many details of the GPU memory hierarchy are not released by GPU vendors. In this paper, we propose a novel fine-grained microbenchmarking approach and apply it to three generations of NVIDIA GPUs, namely Fermi, Kepler, and Maxwell, to expose the previously unknown characteristics of their memory hierarchies. Specifically, we investigate the structures of different GPU cache systems, such as the data cache, the texture cache and the translation look-aside buffer (TLB). We also investigate the throughput and access latency of GPU global memory and shared memory. Our microbenchmark results offer a better understanding of the mysterious GPU memory hierarchy, which will facilitate the software optimization and modelling of GPU architectures. To the best of our knowledge, this is the first study to reveal the cache properties of Kepler and Maxwell GPUs, and the superiority of Maxwell in shared memory performance under bank conflict.