Research Spotlight Archive

Back to Research Spotlight

Title: Green Software: Optimization for Performance, Energy, and Temperature

Participants: Can Hankendi (PhD ’15) and Ashfaquzzaman Khan (PhD ’13); Professors Ayse Coskun (PI) and Martin Herbordt (PI)

Funding: College of Engineering, Dean’s Catalyst Award ’10

Background: Energy efficiency is one of the central societal and technical issues of our time. Today, datacenter energy consumption makes up 2-3% of total U. S. electricity use and increases by 15% every year. The common objective in state-of-the-art energy management techniques is to maintain a desired level of performance while reducing energy consumption. A closely related issue is thermal management – high power consumption not only increases operational energy costs but also causes high temperatures and thus dramatically raises cooling costs. There has been a move to significantly lower the energy costs of cooling by using less expensive infrastructures to replace the HVAC systems used today. Replacing the HVAC systems, however, results in much less predictability and significantly higher temperatures, degrading system reliability and performance.

Considering the complexity of large scale computing infrastructures and time-to-market restrictions, it is not cost-efficient to address these three pressing challenges – Performance, Energy, and Temperature (PET) – solely through novel hardware design. We know that workload has a significant affect on all three parameters. Obviously, better software gets better performance; less obvious is that there is also better software for reducing energy consumption and thermal problems. For example, the distance that data must be routed before it can be operated upon relates directly to the energy required for that operation. There are similar effects for the variance in component switching. Reducing this energy per computation often requires non-obvious software restructuring or introducing non-intuitive “extra” computations.

A major challenge in jointly optimizing for performance, energy, and temperature is the complex interplay among these three metrics. The figure above demonstrates the PET tradeoffs for the parallel benchmarks in the Parsec suite, measured on an Intel I7 server.

Description: The research goal of this proposal is to develop inexpensive, widely applicable methods for generating Green Software for reducing the total cost of computing while achieving high performance and reliability. Specific aims of the project are:

(1) Designing mechanisms for creating software variations that are plausibly optimal with respect to performance, energy, and temperature. Some of these will be based on existing methods of optimizing for performance such as code transformations and autotuning but with additional metrics applied, without adding to development complexity. Case studies will include production applications from document/media processing, scientific computing, and bioinformatics.

(2) Designing consolidation methods to manage resource sharing of software on computing clusters for enabling energy proportional computing. The methods will include explicit consideration of the temperature effects and cooling as well as the equipment and energy costs.

Recent Publications:


Pages: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36