|M.Sc Student||Lior Zeno|
|Subject||I/O-Intensive Workloads on Accelerators|
|Department||Department of Electrical Engineering||Supervisor||Professor Silberstein Mark|
|Full Thesis text|
In this thesis, we focus on two I/O-intensive workloads on accelerators: (a) I/O-driven preemption on GPUs, and (b) Streaming applications on SmartNICs.
Recent studies have shown the benefits of native GPU I/O layers, in terms of both programmability and performance. However, the GPU threads performing I/O calls are forced to busy-wait for the completion of I/O operations, resulting in underutilized hardware, higher power consumption, and reduced system throughput. We argue that I/O-driven preemption improves the performance of existing solutions, despite many challenging system characteristics such as a large kernel state. We present GPUpIO, a software-based I/O-driven mechanism for GPUs. We show that our prototype may double the effective system throughput by completely hiding the I/O latency behind computations, and the limitations of software-based solutions.
With rising datacenter network rates, cloud vendors are deploying FPGA-based SmartNICs to achieve cost-effective acceleration for hypervisor networking tasks and network functions. However, due to (a) lack of programming models and OS abstractions, and (b) limited FPGA resources and the programmability wall, attempts to use SmartNICs’ inline data processing capabilities for accelerating general purpose server application in clouds have been restricted. We describe the challenges and design considerations for accelerating general-purpose streaming applications on SmartNICs, and demonstrate the compelling performance SmartNICs achieve for several real-world workloads such as an Node.js IoT server and the memcached key-value store .