Efficient Resource Management
Data centers have become an important factor in our everyday life: Internet services run on their servers, researchers buy processing time at short notice for time-consuming calculations, and even the end-user makes more and more use of applications that run in the cloud. This rapidly growing demand for short-termed, guaranteed and well-tailored computational power takes current system architectures to their limits. One way to face these problems are so called heterogeneous architectures: parallel systems which not only contain several (possibly different) CPU cores but also accelerator cores (e.g., GPGPUs or FPGAs). Such accelerators may enable data centers to schedule time-critical jobs without disturbing normal operations. Moreover, the possibility to migrate parts of algorithms to specialized hardware – perhaps even during runtime – may enable largely improved solution strategies. Apart from classical optimization objectives, such as time or space, we also consider more advanced concepts (e.g., energy and temperature).
Our research focuses on algorithmic problems that occur in data centers made up of such heterogeneous systems. How can such a large amount of different resources be managed efficiently? Additional questions include the incorporation of concepts like profitability and support for service level agreements. We develop theoretical models and provably good strategies, with the goal of creating a deeper understanding of the problem at hand. Initial models consider a number of speed-scalable processors whose speed can be adjusted at runtime. A higher speed processes customers’ jobs faster, possibly achieving higher revenue, but causes higher energy costs. Moreover, missed deadlines may cause a loss in profit or even the need for compensation. While such models do not admit optimal strategies, good strategies should take these and similar consequences into account and try to achieve at least a (proven) virtually optimal schedule.
Service level agreements modeled as functions
Distributed Computing in Networks
Volunteer computing represents a cost effective alternative to conventional data centers. The general idea is to exploit the left-over computational power of desktop computers distributed all over the world. Candidates for such systems are standard PCs in many companies, since most of their uptime is spent in idle mode. Managing such a network of volunteering computers is a challenging task: the available computational capacity can vary significantly depending on the location, the user behavior and the current time. Moreover, computers may simply be switched off or crash. Our approach to this problem is a peer-to-peer based middleware called PUB-Web (Paderborn University BSP-based Web Computing). Its peer-to-peer based design seems well suited for such a highly dynamic environment and allows itself to adopt quickly to the current system state. Intelligent load balancing strategies distribute the various parallel processes over the available PCs. This is done during runtime and in consideration of the corresponding system load of the various computers. Regularly and redundantly stored backups of a process state allows our system to deal with unexpected failures of individual PCs.
Our current research focuses on the possibilities to expand PUB-Web beyond its current field of application. We plan to use it as a basis for a network architecture interconnecting both traditional desktop systems and groups of data centers. Participants in the network may provide portions of their computational power, possibly on a contractual basis.
Different levels of scheduling problems
Contact:
Dipl.-Math. Peter Kling, M.Sc.
E-mail: Peter.Kling@hni.uni-paderborn.de
Phone: +49 (0) 5251/60 64 27
