[cs-talks] Upcoming Seminar: Christine Bassem, PhD Defense (Wednesday, June 24)
fgreen1 at bu.edu
Tue Jun 16 14:45:58 EDT 2015
Route Coordination in Brokered Environments
Wednesday, June 24, 2015 at 10am in MCS 180 (Hariri Institute)
Abstract: The increased use of federated commodities to provide scalable, high quality services is fueling research in novel brokerage mechanisms and models for the management of resources under the control of rational agents. In such settings, the responsibility of a broker varies from coordinating the communication between agents to making decisions on their behalf. This thesis focuses on brokerage services that require routing of -- as well as route coordination for -- a set of commodities over a communication/mobility graph. Routing can either be explicit or implicit. In explicit routing, the location of the commodities, and hence the endpoints for the communication, are fixed and the problem reduces to the identification of a path subject to some objectives and constraints. In implicit routing, the routes are fixed, and the problem reduces to the assignment of endpoint locations to the commodities subject to some objectives and constraints.
In this talk, I will overview the thesis contribution, focusing on a model of implicit route coordination in a cloud setting. In particular, I will present results related to providing execution predictability guarantees for data-intensive workloads in a federated cloud environment. I start with a definition of the problem of packing virtual machines (VMs) in network-constrained physical hosts with an objective to minimize the cost of communication among these VMs, and show it to be as hard as the Bin Packing problem. Following the definition of a heuristic to solve the packing problem for a single workload, I present the problem of scheduling a batch of data-intensive workloads with flexible deadlines, and a batch scheduling mechanism that accounts for each workload's network demands as well as its time-flexibility. Then, I present a VickreyClarkeGroves-based pricing mechanism that fairly distributes the broker's cost over the different workloads according to their contribution to that cost. Finally, I'll extend these mechanisms to the scheduling of data-intensive workloads in online (as opposed to batch) settings.
Ibrahim Matta (reader)
Mark Crovella (chair)
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks