Mark what times you are available, hit the Edit link on top. After editing, scroll down and hit save.
9:15am - 10:00am: Karu Sankaralingam (6381CS)
10:00am - 10:45am: Jignesh Patel (4357 CS)
10:45am - 11:00 talk setup
11:00am - noon: Talk in 4310 CS
noon - 1:30pm: Lunch (Shan, Mike)
1:30pm - 2:15pm: Nam Kim (4615 Engineering Hall)
2:30pm - 3:00pm: Mike Swift (7369 CS)
3:00pm - 3:30pm: Shan Lu (7367 CS)
3:30pm - 4:00pm: Charles Fischer (6367 CS)
4:00pm - 5:30pm: Architecture students (4310 CS)
7:00pm: Dinner (Karu Sankaralingam,Ben Liblit, David Wood)
10:00am - 10:45am: David Wood (6369 CS)
10:45am - 11:30: Mark Hill (6373 CS)
11:30am - 1:00: Lunch with PL students
1:30pm - 2:15pm: Ben Liblit (6357 CS)
2:15pm - 3:00pm: Aditya Akella
3:00pm: Leave for airport
Energy-Efficient Software and Hardware for Data Centers
Companies such as Amazon, Google, Microsoft, and Yahoo are building several large data centers containing tens of thousands of machines to provide the computational capability to support a variety of web services such as search, email, and online shopping. Energy costs for operating and cooling these machines are a large fraction of the monthly operational costs of running these facilities. We discuss opportunities for lowering this cost by improving the energy efficiency of data center software and hardware used to run an industry-strength web search engine.
On the software side, we propose a system called Green that provides a simple and flexible framework that allows programmers to take advantage of application-specific Quality-of-Service (QoS) constraints to perform approximations that reduce the amount of work that needs to be performed, and consequently energy consumed, while providing statistical guarantees that the QoS constraints are met. We used Green to reduce the energy consumption of web search by 18% with QoS degradation of 0.27%.
On the hardware side, we investigate the possibility of using small mobile-class architectures that implement simpler data paths to conserve power and, thereby, deliver throughput at a much lower power cost. Since data center applications operate under strict QoS and latency constraints, we evaluate the impact of small cores on an application’s computational intensity and micro-architectural activity. We show that web search on a mobile-class architecture, such as Intel’s Atom processor, is 5 times more efficient that on a server-class architecture, such as Intel’s Xeon processor, but we must mitigate the price of efficiency with a coordinated strategy that might involve over-provisioning, micro-architectural enhancement and accelerators.
Bio: Trishul Chilimbi is a senior researcher at Microsoft Research leading the Runtime Analysis & Design (RAD) research group, which is part of Microsoft’s Research in Software Engineering (RiSE) organization. He received his Ph.D in Computer Science from the University of Wisconsin at Madison in 1999 and joined Microsoft Research immediately thereafter. His areas of interest are programming languages, compilers, runtime systems, computer architecture, and parallel and distributed systems. He is currently focused on improving the performance and energy-efficiency of web services both from a client and data center perspective.