1. Green Computing for a Clean Tomorrow Improve efficiency, reliability, availability, and usability of computing systems.  Sacrifice a bit of raw speed to reduce power & energy consumption.  Improve overall throughput as the system will always be available, i.e., effectively no downtime. Reduce total cost of ownership & increase return on investment. Crude Analogy Formula One Race Car: Wins raw performance but reliability is so poor that it requires frequent maintenance. Throughput low. Honda S2000: Loses raw performance but high reliability results in high throughput (i.e., miles driven/month  answers/month). 1.Power Consumption & Heat Generation Hurt Reliability, Availability, & Total Cost of Ownership 2.Electrical Power for Computing Costs $$$ Earth Simulator: 12 MW/year  $10M/year World’s Processors: 1.3 GW/year  $1B/year “Hiding in Plain Sight, Google Seeks More Power,” The New York Times, June 14, 2006. 3.Computing “Contributes” to Global Warming Motivation Goal: Deliver high performance while reducing power & energy consumption and improving reliability. Prof. Wu FENG, feng@cs.vt.edu College of Engineering, Depts. of CS & ECE Just the processors (i.e., CPUs) in PCs ≈ 40 Hoover Dams (Estimated power consumption of PCs ≈ 120 Hoover Dams) 2001 to 2006 1 10 100 1000 1.5  11 0.7  0.5  0.35  0.25  0.18  0.13  0.1  0.07  I386 – 1 watt I486 – 2 watts Pentium – 14 watts Pentium Pro – 30 watts Pentium II – 35 watts Pentium III – 35 watts Chip Maximum Power in watts/cm 2 Surpassed Heating Plate Not too long to reach Nuclear Reactor Year Pentium 4 – 75 watts 1985 1995 2001 Itanium – 130 watts Source: Fred Pollack, Intel. New Microprocessor Challenges in the Coming Generations of CMOS Technologies, MICRO32 and Transmeta SystemsProcessorsReliability & Availability ASCI Q8,192MTBI: 6.5 hrs. 114 unplanned outages/month. –Outage sources: storage, CPU, memory. ASCI White 8,192MTBF: 5 hrs. (2001) and 40 hrs. (2003). –Outage sources: storage, CPU, 3 rd -party HW. NERSC Seaborg 6,656MTBI: 14 days. MTTR: 3.3 hrs. – SW is the main outage source. Availability: 98.74%. PSC Lemieux 3,016MTBI: 9.7 hrs. Availability: 98.33%. Google~450,000550+ reboots/day; 2-3% machines replaced/yr. – Outage sources: storage, memory. Availability: ~100%. MTBI: mean time between interrupt; MTBF: mean time between failure; MTTR: mean time to restore “Making a Case for a Green500 List,” 20 th IEEE Int’l Parallel & Distributed Processing Symp., Apr. 2006. “A Power-Aware Run-Time System for High-Performance Computing,” SC|05, Nov. 2005. “The Importance of Being Low Power in High-Performance Computing,” CTWatch Quarterly (NSF), 1(3):12-20, Aug. 2005. “Green Destiny and Its Evolving Parts,” Innovative Supercomputer Architecture Award, 19 th Int’l Supercomputer Conf., Jun. 2004. “Green Destiny + mpiBLAST = Bioinfomagic,” 10 th Int’l Conf. on Parallel Computing (ParCo), Sept. 2003. “Honey, I Shrunk the Beowulf!” 31 st Int’l Conf. on Parallel Processing, Aug. 2002. Self-Adapting Software for Energy Efficiency Conserve power & energy WHILE programs run. Sequential Codes Parallel Codes Energy savings and performance improvement! relative time / relative energy with respect to total execution time and system energy usage 1.Low-Power, High-Performance Computing  Green Destiny : A 240-Node Supercomputer in 5 Sq. Ft. with a 3.2-kW Power Envelope  Reliability Operating Environment: A dusty 85°-90° F warehouse. No machine room. No unscheduled downtime in its 24-month lifetime.  Performance: A Top500 Supercomputer (circa March 2002). Linpack: 101 Gflops. 4 40 Arrenhius Equation (applied to microelectronics) & Twenty Years of Empirical Data For every 10°C increase in temperature, the failure rate of the system doubles. Reduce power consumption  Reduce system temperature  Reduce failure rate Observation The Project: Supercomputing in Small Spaces Green Destiny : A 240-node Supercomputer in Five Sq. Ft. 2.Power-Aware, High-Performance Computing Only Difference? The Processors Green Destiny: Low-Power Supercomputer Green Destiny “Replica”: Traditional Supercomputer 3.2 kW 30.0 kW Hypothesis Selected Publications  Many commodity technologies support dynamic voltage & frequency scaling (DVFS), which allows changes to the processor voltage and frequency at run time.  A computing system can trade off processor performance for power reduction. Power  V 2 f, where V is the supply voltage of the processor and f is its frequency. Processor performance  frequency.  Approach: Intelligent DVFS Scheduling  Determine when to adjust the voltage-frequency setting and what to adjust it to. Approaches Results Laboratory NAS/NPB 3.2 – MPI, C.16 Featured in The New York Times, CNN, and BBC News Now in The Computer History Museum