Semantic-less Coordination of Power Management and Application Performance HOTPOWER ‘09
About the Author(1) Aman Kansal Microsoft Research Redmond, WA
About the Author(2) Jie Liu Microsoft Research Redmond, WA
Abstract Different power management modules affect each other. Try to find out an approach for semantic-free coordination. Design an interface at the system and application layers.
Semantic-less implies… The values shared via the interface cannot be compared to other values. A module only know it’s own values, don’t know whether a higher or lower value is better. The goal is to compose multiple modules, with their independent strategies.
Contributions A semantic-less mechanism: a narrow data interface and a generic coordination algorithm. Applicaton tunes QoS, System changes P-state(processor voltage and frequency) without konwing anything about the other.
Related works The coordination among system modules[6], application[3], and applications and system modules[11,12,8,2]. Joint optimizations of system and application performance[1,7]. All methods assume semantic information and the coordinated entities.
Power Performance Coordination
Coordinated system design(1)
Coordination interface App(i) publishes QoS(i) System modules publishes P(j), other modules don’t know what it means.(P-state, throughput cap, sleep mode, etc) System also publishes a signal C in {-1, 0, 1} to indicate if energy needs to be reduced(c = -1) or keep constant(c = 0) or is available for increasing(c = 1). (A system power measurement or estimate derived from performance counters based power models[5, 10] may be help determine whether power usage need to reduce)
Coordination Algorithm(1) System algorithm
Application Algorithm(1) If no module acquire the lock and the system determines hat energy usage must be reduced then is sets C = -1. The action for C =1 is similar.
Application Algorithm(2) After this, the system will detect at least one QoS(i) changed. The system then change i’s state. If the resource utilization reduced to a desired value and the system updates C = 0. It will not supply C = 1 if previous configuration did not have the target power usage to prevent oscillations.
Experiment 1 Consider a battery operated laptop decoding high definition video. (on a Lenovo T61p)
Experiment 2(1) This experiment considers multiple applications with different functionalities sharing a server. A stream server serves HD (3.2Mbps), DVD(2Mbps), broadband (300kbps), dial-up(28 kbps. Suppose the revenues (QoS levels) are $4, $3, $2, $0.5. Varying CPU usage: 100%, 75%, 50% and 25%. The conversion of searches to purchases varies with search quality leading to varying revenues of $6, $5, $4, and $1 respectively. On an HPDLG380 blade server with 2x4-core Xeon processors and an 8-disk RAID array.
Experiment 2(2)
Reference [1] P. Bodik, R. Griffith, C. Sutton, A. Fox, M. Jordan, and D. Patterson. Automatic exploration of datacenter performance regimes. In First Workshop on Automated Control for Datacenters and Clouds (ACDC09), Barcelona, Spain, June [2] S. Brandt, G. Nutt, T. Berk, and J. Mankovich. A dynamic quality of service middleware agent for mediating application resource usage. In IEEE Real-Time Systems Symposium, December [3] J. S. Chase, D. C. Anderson, P. N. Thakar, A. M. Vahdat, and R. P. Doyle. Managing energy and server resources in hosting centers. SIGOPS Oper. Syst. Rev., 35(5):103–116, [4] D. D. Corkill. An introduction to blackboard systems. AI Expert, 6(9):40–47, September [5] X. Fan, W.-D. Weber, and L. A. Barroso. Power provisioning for a warehouse-sized computer. In Proceedings of the International Symposium on Computer Architecture (ISCA), June [6] J. Heo, D. Henriksson, X. Liu, and T. Abdelzaher. Integrating adaptive components: An emerging challenge in performanceadaptive systems and a server farm case-study. In IEEE International Real-Time Systems Symposium, pages 227–238, [7] X. Liu, P. Shenoy, and M. D. Corner. Chameleon: Application Level Power Management. IEEE Transactions on Mobile Computing, 7(8):995–1010, August [8] R. Nathuji, P. England, P. Sharma, and A. Singh. Feedback driven qos-aware power budgeting for virtualized servers. In Feedback Control Implementation nd Design in Computing Systems and Networks (FeBID), San Francisco, CA, USA, April [9] R. Raghavendra, P. Ranganathan, V. Talwar, Z. Wang, and X. Zhu. No ”power” struggles: coordinated multi-level power management for the data center. SIGOPS Oper. Syst. Rev., 42(2):48–59, [10] S. Rivoire, P. Ranganathan, and C. Kozyrakis. A comparison of high-level full-system power models. In HotPower’08: Workshopon Power Aware Computing and Systems, December [11] X.Wang and Y.Wang. Co-con: Coordinated control of power and application performance for virtualized server clusters. In 17 th IEEE International Workshop on Quality of Service (IWQoS), Charleston, South Carolina, July [12] W. Yuan and K. Nahrstedt. Energy-efficient soft real-time cpu scheduling for mobile multimedia systems. In SOSP, October 2003.