Home > Seminars > Resource Allocation in Datacenters

Resource Allocation in Datacenters


3/20/2018 at 11:00AM


3/20/2018 at 12:00PM


258 Fitzpatrick Hall


College of Engineering close button

Thomas Fuja

Thomas Fuja

VIEW FULL PROFILE Email: tfuja@nd.edu
Phone: 574-631-7244
Office: 275 Fitzpatrick Hall


Wireless Institute Professor
Prof. Fuja research addresses reliable communication over inherently unreliable and/or constrained communication links. He has recently focused his research on the changing role that channel codes play in the context of wireless networks, i.e., to not only provide physical-layer robustness but ...
Click for more information about Thomas
Add to calendar:
iCal vCal

With the growing dependence upon datacenters to deliver high-quality services, efficient resource allocation in datacenters becomes increasingly important. Exciting opportunities for system redesign, as well as algorithmic/theoretical challenges unique to datacenters, arise across all layers of the system. In this talk, I will present my work that features two complementary approaches to provably efficient resource allocation.

In the first part, I will discuss the problem of scheduling data-intensive applications. A fundamental problem is data locality --- tasks consume different amounts of resources at different locations. We formulate this problem as scheduling in a multi-class parallel-server system. Previous work by Harrison, Williams and Stolyar has left open the problem of achieving delay optimality with unknown arrival rates. We present a novel class of algorithms that achieve precisely this goal, and develop new proof techniques are to establish its delay optimality. Experiments on Amazon EC2 show >10x acceleration over existing schedulers.

In the second part, I will present a complementary, data-driven approach to resource allocation that aims at solving complicated, analytically intractable problems. Many such problems can be modeled by a general class of Markov Decision Processes with a continuous state space. We develop the Nearest Neighbor Q-Learning algorithm, which learns the optimal policy directly from observations of the system dynamics. We provide finite-sample, polynomial-time guarantees on the performance of our algorithm in an online setting.

Seminar Speaker:

Dr. Qiaomin Xie

Dr. Qiaomin Xie

Massachusetts Institute of Technology

Qiaomin Xie is a postdoctoral researcher with the Laboratory for Information and Decision Systems at MIT. In Fall 2016, she was a research fellow at the Simons Institute for the Theory of Computing. Qiaomin received her Ph.D. in Electrical and Computing Engineering from University of Illinois Urbana Champaign in 2016, and graduated from Tsinghua University with a B.E. in Electronic Engineering. Her research interests lie in the broad area of computer and networked systems, with a recent focus on resource allocation in datacenters as well as learning-based networked systems. She is the recipient of UIUC CSL PhD Thesis Award (2017), the Yi-Min Wang and Pi-Yu Chung Research Award (2015), and the best paper award from IFIP Performance Conference (2011).