Oblivious and cost aware load sharing in the Cloud: Oblivious and cost aware load sharing in the Cloud

Lecture: Departmental | November 12 | 4-5 p.m. | Soda Hall, Wozniak Lounge (Room 430-438)

Speaker/Performer: Danny Raz, Technion, Israel

Sponsor: Electrical Engineering and Computer Sciences (EECS)

In many cases, large scale cloud-based services are provided simultaneously from several, potentially distant sites. The actual choice of the specific site and the specific server that would fulfill a given user request has a critical impact on the overall performance of the service. This gives rise to a highly complex optimization problem, which often involves multiple objectives and many parameters. Irrespective of the precise optimization criteria, any attempt to address such an optimization problem will incur significant overhead by collecting the required (state-dependent) information from the various network locations. One way to address this problem is through an oblivious approach, i.e., a distributed load-sharing scheme that does not use any state information. We revisit this extensively studied problem and present a novel scheme, based on creating, in addition to the regular job requests that are assigned to a randomly chosen server, also low priority job request replicas that are sent to a different randomly chosen server. We show that, when servers can coordinate the removal of redundant copies upon completion of a job, the performance of the system exhibits dramatic improvement of up to 45% even under high load conditions. When no such coordination is possible, a simple timeout mechanism yields a significant improvement of up to 15%.

Another way to address this challenge is to use the server state information, while considering the cost of obtaining the information in the scheme. All previous models have not incorporated this monitoring cost within them. Our focus is on a rigorous study of the right amount of monitoring, that is, we want to maximize system utility by monitoring the needed servers without over monitoring. Following the theoretical model we develop several practical approaches in this context and study their expected performance.

This talk is based on joint papers with David Breitgand, Rami Cohen, Amir Nahir, and Ariel Orda.

Event Contact: seanmc@eecs.berkeley.edu, 510-643-0264