ECE PhD Dissertation Defense: Mahsa Bayati

PhD Defense: Efficient Data Access with Heterogeneous Computing using GPUs and Direct Non-volatile Storage.

Mahsa Bayati

Location: Zoom Link

Abstract: The amount of data being collected that requires analysis is growing at an exponential rate. Along with this growth comes an increasing need for storage and computation. Researchers address these needs by (I) deploying distributed bigdata platforms equipped with cutting-edge storage devices, and (II) building heterogeneous clusters with Central Processing Units (CPUs) and computational accelerators such as Graphics Processing Units (GPUs). The high performance of these mainstream systems is achieved by efficiently accessing data and computation resources and scheduling parallel and distributed tasks.

The performance of each job depends on the characteristics of both the application and the underlying storage and computational environments. However, it is not a trivia to maintain efficiency and provide high performance in these mainstream systems. First, in bigdata platforms like Spark and Hadoop, full utilization of Solid State Devices (SSDs), i.e., Non-Volatile Memory Express (NVMe) and Key-Value (KV) SSDs is challenging. Data communication between Spark tasks, levels of parallelism, and resource co-location significantly affects achieving high I/O throughput. Second, in heterogeneous systems, one of the main bottlenecks of GPU computation is the data transfer bandwidth to GPUs in I/O intensive applications. The traditional GPU approach gets data from host memory, which can limit data throughput and processing and thus degrade end-to-end performance. In this work, we initially explore different attributes to exploit the full benefits of various SSDs in bigdata platforms. We then focus on mitigating the data transfer bottleneck in a heterogeneous bigdata framework.

We build a heterogeneous framework that facilitates GPU direct access to storage. Our framework aims to minimize the data transfer delay, thus enhancing the performance of distributed and parallel tasks to obtain the full benefits of compute and storage resources. Our heterogeneous cluster is supplied with CPUs and GPUs as computing resources and non-volatile flash-based drives as storage resources. We also deploy the Spark bigdata platform to execute large workloads over our cluster. We then adopt a novel technique (e.g., Peer-to-Peer Direct Memory Access) to connect GPUs to non-volatile storage directly. Experimental results reveal that our heterogeneous Spark platform successfully bypasses the host memory and enables GPUs to communicate directly to the NVMe drive, thus achieving higher data transfer throughput. The contributions of the dissertation are: (I) Realizing that bigdata processing applications need to consider framework features and application characteristics to fully utilize the high bandwidth of modern SSDs, where compute and storage locality is essential to optimize the cost and performance. (II) Deploying our novel heterogeneous framework supporting GPU direct storage access improves data communication time around 35%- 50% and end-to-end performance by 30%.