Department: Computer Science & Engineering
Faculty Advisor(s): Michael Taylor
Award(s): Department Best Poster

Primary Student
Name: Anshuman Gupta
Email: a2gupta@ucsd.edu
Phone: 858-534-6887
Grad Year: 2013

Cloud Computing aims to provide a cheap, highly fluid, nearly limitless, on-demand pool of computational resources made available via the Internet. Some of the cloud service providers use multiprogrammed data centers, and they expose these resources using the abstraction of virtual machines. On the other hand, hardware vendors have demonstrated manycore chips that scale to a hundred or more cores, and postulated the use of these chips in the context of these large manycore multiprogrammed data centers. As a result, we might soon see the emergence of manycore multiprogrammed data centers. These manycore chips rely on shared resources, such as cache and memory bandwidth; as a result, performance is especially sensitive to the interference between competing applications in a multiprogrammed environment. The resulting unpredictable performance makes it harder for the data center operators to make high-level decisions such as resource allocation. We propose a manycore architecture augmented with novel hardware mechanisms for reducing resource interference across applications. The results provide a compelling argument for adding the requisite micro-architectural structures in commercial manycore chips targeted towards multiprogrammed data centers.

« Back to Posters or Search Results