Presentation on theme: "Published on: HotPower ‘09 Presentation By: Liang Hao"— Presentation transcript:
1 Some Joules Are More Precious Than Others: Managing Renewable Energy in the Datacenter Published on: HotPower ‘09Presentation By: Liang HaoTuesday, August 03, 2010
2 Author Christopher Stewart from The Ohio State University Kai Shen from University of RochesterPublications:Of Christopher StewartOf Kai Shen
3 MotivationTo reduce datacenters’ dependence on costly and less clean energy from the gridHence to maximize the use of renewable energiesTo explore the possibility of evaluating the request-level power consumption
4 Problems Datacenters powered by renewable energy need backups Applications in the datacenter must be available 24x7But wind and solar energy are intermittentDatacenters powered by renewable energy need backupsPrimary options: Grid, generator, batteryAlternatives are either dirty and/or costlyRenewables are precious!renewable = joule converted from solar/windThe preferred energy sourceAvailable only sometimes and costly to store
5 Opportunities Capacity planning Load balancing Compute power should fluctuate with intermittent outages—i.e., turn machines offLoad balancingRoute requests to datacenters with unused renewablesMigrate services to datacenters with renewables
7 Intermittency 1.Datacenter Modeling Automatic transfer switch (ATS)Input 2 power sources, outputs 1 power sourceMonitors power from the primary sourceWhen power from primary dips below threshold, ATS switches to secondaryWhen primary exhibits power of threshold, ATS switches back to secondary
8 Intermittency 1.Datacenter Modeling Key parameters related to the ATSto ensure dependability or reliability, threshold equals peak consumptionto make full use of renewable energies, scale down the threshold
9 Intermittency 2.Wind Intermittency Battery backup too costlyBut if we apply renewable-aware management, there is enough supply from other datacenters
10 Intermittency 3.Renewable Utility when threshold set to peak power, the utility of wind turbine power production drops 65% compared to when threshold set to zero.
11 Intermittency 3.Renewable Utility Economical feasibility (metric: cost per KW-hour)Average price for commercial electricity $0.10 KW-hour$2.4M to erect a wind turbine that is connected (directly) to a datacenter [European Wind Energy Assc.]$1.6M installation2% annual maintenance feesLifetime of turbine: 20 yearsDatacenter at CA or MT could use 24M KWhEither high power consumption or zero thresholdWind-powered datacenter in MT: $0.04 KWh
12 Request-level Event Profiling To estimate the power consumption of individual requestsWith quantized statistics, scheduler could possibly route some requests to datacenters with unused renewables
13 Request-level Event Profiling Tracing the route that a request go through, including CPU usage and other hardware events.We configured the performance counters to assemble three predictor metrics for our power model:L2 cache requests per CPU cycle (Ccache),memory transactions per CPU cycle (Cmem),and the ratio of non-halt CPU cycles (Cnonhalt).
14 Request-level Event Profiling Power consumption is calculated according to the expression belowwhere P’s are coefficient parameters for the linear model are constants that approximate ceiling values for the predictor metrics.
15 Request-level Event Profiling micro benchmarks 1) idle2) CPU spinning with no access to cache or memory3/4) Apache web server with either short requests (no more than 1KB files) or long requests (files of 100 KB– 1 MB)5/6) OpenSSL RSA encryption/decryption using either a small key or a large key. We also use four full server workloads7) TPC-C running on the MySQL database8) TPC-H running on the MySQL database9) RUBiS10) WeBWorK .
16 Request-level Event Profiling Request workloads executed in isolationWattsUp power meter measures watts and joulesProcessor was not adjusted during tests
18 Reference Green House Data: Greening the data center. Realistic nonstationary workloads. Wind power. Google solar panel project.June 2007. P. Barham, A. Donnelly, R. Isaacs, and R. Mortier.Using Magpie for request extraction and workload modeling.In USENIX Symp. on Operating Systems Design and Implementation, Dec F. Bellosa. The benefits of event-driven energy accounting in power-sensitive systems.In 9th ACM SIGOPS European Workshop, Sept J. Chase, D. Anderson, P. Thakar, A. Vahdat, and R. Doyle.Managing energy and server resources in hosting centers.In ACM Symp. on Operating Systems Principles, Oct European Wind Energy Association. The economics of wind energy. U. Hölzle. Powering a Google search.Jan
19  D. Meisner, B. Gold, and T. Wenisch. Powernap: Eliminating server idle power.In Int’l Conf. on Architectural Support for Programming Languages and Operating Systems, Mar National Renewable Energy Laboratory. NREL: Western wind resources dataset K. Shen, M. Zhong, S. Dwarkadas, C. Li, C. Stewart, and X. Zhang.Hardware counter driven on-the-fly request signatures.In Int’l Conf. on Architectural Support for Programming Languages and Operating Systems, Mar C. Stewart, T. Kelly, and A. Zhang.Exploiting nonstationarity for performance prediction.In EuroSys Conf., Mar C. Stewart, M. Leventi, and K. Shen.Empirical examination of a collaborative web application.In IEEE Int’l Symp. On Workload Characterization, Seattle, WA, Sept Benchmark available at P. Thibodeau. Wind power data center project planned in urban area. ComputerWorld, Apr A. Vahdat, A. Lebeck, and C. Ellis.Every joule is precious: the case for revisiting operating system design for energy efficiency.In ACM SIGOPS European Workshop, Sept