Dec 8, 2017
Lopata Hall, Room 101
"Stochastic Configuration Tool for Heterogeneous Computing Applications"
Adviser: Roger Chamberlain
In the fight to find the next big thing in computing many are turning to field programmable gate arrays (FPGAs) in heterogeneous computing as a source of hardware acceleration for their versatility, power consumption, and potential computation speed. Recently, a system designed by Intel and their newly acquired partner, Altera, hopes to bring the latest and greatest in heterogeneous computing in the form of the Hardware Accelerator Research Program (HARP). The HARP system implements a closely interconnected FPGA and CPU in an attempt to utilize the strengths of both platforms. Programing the customizable hardware on the HARP can be done using high level synthesis (HLS) tools such as the Intel FPGA SDK for OpenCL. When using HLS tools there is the potential for a large number of configurations that may be functionally correct but not the optimal solution for the hardware or algorithm implementation. To this end, we are investigating the use of a stochastic optimizer in an attempt to find the optimal configuration given the execution model or simulation results. However, there is a potential for the models of evaluation to be inaccurate. We are investigating the impact of model error on the stochastic optimization process.
"Towards a More Efficient Cloud Based Computational Infrastructure"
Adviser: Roch Guérin
Infrastructure-as-a-service (IaaS) is the most popular service among different kinds of cloud services. The main part of IaaS is to give users on-demand access to the resources. To guarantee the on-demand service, the service provider either over-provides the service, or rejects a large amount of user requests. Meanwhile, not all users are expecting an on-demand service. Many applications, like high-throughput computing (HTC), can use any available resources and terminate when the resource is no longer available.
We propose a cloud infrastructure system combining reserved and on-demand allocation of resources, as well as an opportunistic provisioning of cycles from unused nodes and idle reserved nodes. For experimental evaluation, we extend the OpenStack cloud computing toolkit to deploy backfill virtual machines on idle cloud nodes, as well as "borrow" resources from idle reserved machines for HTC workload.