Machine learning (ML) is now omnipresent in all spheres of life. The use of deep neural networks (DNNs) for ML has gained popularity over the past few years. This is because DNNs are capable of efficiently solving complex problems such as image processing, object detection, language processing, etc. To train these DNN workloads, graphics process- ing units (GPUs) have become the most widely used platform. A GPU can support a large number of parallel threads that execute simultaneously to achieve a very high throughput. However, as the sizes of the DNN workloads grow, a single GPU is no longer adequate to provide fast training, and developers resort to using multi-GPU (MGPU) systems that can reduce the training time significantly. Consequently, to keep pace with the growth of DNN applications, GPU vendors are actively developing novel and efficient MGPU systems.
To better understand the challenges associated with designing MGPU systems for DNN workloads, in this thesis, we first present our efforts to understand the behavior of the DNN workloads, in particular, the training of DNN workloads on MGPU systems. Using the DNN workloads as benchmarks, we observe the evolution of MGPU system architecture. Based on our profiling and characterization of DNN workloads on existing high-performance MGPU systems, we identify the computation- and communication- intensiveness of the DNN workloads and the hardware- and software-level inefficiencies present in the existing MGPU systems. We find that the data movement across multiple GPUs and high remote data access cost leading to NUMA effects, data duplication, and inefficient use of GPU memory leading to memory capacity issues, and the complexity in programming MGPUs pose serious limitations in the execution of ever-scaling DNN workloads on MGPU systems.
To overcome the limitations of existing MGPU systems, we propose to unify the main memory of GPUs to design an MGPU system with true shared memory (MGPU-TSM). Our proposed MGPU-TSM system demonstrates a significant performance boost (3.8× for a 4 GPU system) over the best-performing existing MGPU system. This is because MGPU-TSM system eliminates the NUMA effects and the necessity for data duplication. To provide seamless data sharing across multiple GPUs and ease programming of MGPU- TSM, we propose a light-weight coherence protocol called MGCC. MGCC is a timestamp- based protocol that provides both intra- and inter-GPU coherence. We implement a number of hardware features including unified memory controller, request tracker and timestamp storage unit to support MGCC. Using both standard and synthetic stress benchmarks, we evaluate the MGPU-TSM system with MGCC leveraging sequential as well as relaxed consistency. Our evaluation of a 4-GPU system using MGPUSim simulator suggests that our proposed coherent MGPU system achieves up to 3.8× improved performance than current best-performing MGPU system while the stress tests performed using synthetic benchmarks suggests that MGCC leads to up to 46.1% performance overhead.
| Identifer | oai:union.ndltd.org:bu.edu/oai:open.bu.edu:2144/42590 |
| Date | 15 May 2021 |
| Creators | Mojumder, Md Saiful Arefin |
| Contributors | Joshi, Ajay |
| Source Sets | Boston University |
| Language | en_US |
| Detected Language | English |
| Type | Thesis/Dissertation |
| Rights | Attribution 4.0 International, http://creativecommons.org/licenses/by/4.0/ |
Page generated in 0.0021 seconds