1. Optimal Fan Speed Control for Thermal Management of Servers UMass-Amherst Green Computing Seminar September 21 st, 2009

2. HP Confide ntial 2 Problem Power consumption key challenge in data centers 1.5% U.S. consumption in 2006 (60b kwh) Every 1W server power  0.5-1W cooling power Increasing density == Need more cooling power >76% data centers use blades (and rising) Fan power can be 2000W or 23% in blade enclosures Is airflow overprovisioned? Can we decrease consumption by regulating fan speed better?

3. HP Confide ntial 3 Blade Enclosure Overview 10 fans total: 5 on top, 5 on bottom Fans pull air through blades and out 5 on top 5 on bottom

4. HP Confide ntial 4 Basic Approach Model the cooling/fan system −Need model relating blade temperatures to fan speed −Complex interactions with multiple variables Develop smart fan speed controller −Use model as input to MIMO controller −Compare with existing integrated controller Problem Definition −Minimize fan power consumption −Ensure blade temperatures don’t exceed threshold

5. HP Confide ntial 5 The Models Power models are (kind of) easy Blade power consumption  linear Fan power consumption  cubic Temperature models are hard Complex zonal variations Lack of sensors Steady-state versus transient

6. HP Confide ntial 6 Temperature Models Steady-state: use thermal resistance co-efficient R −CPU transfers heat to ambient air −Relationship between R and temp. change and power consumption −Relationship between R and fan speed −Solve for CPU temp. Transient: use energy balance equation −Takes some time to transfer heat to air −Model as a first-order discrete time system −Validated the models though experiments −Models have 167 parameters!!! −Basically, 10 blades with 16 fans

7. HP Confide ntial 7 Fan Controller Optimization problem is hard −Fan speed to power consumption relationship is non-linear −Transient model has non-linear relationship b/t CPU and fan speed −May not have feasible solution Simplify the problem −Find minimum airflow needed at each blade (locally) −Solve the global problem using these airflows −Compare with Integrated Fan Controller −If temp goes up speed fan up −If temp goes down lower fan speed

8. HP Confide ntial 8 Results Reduced fan power from 213W to 172W or 20% (Figure 9) On-demand cooling of high utilization CPUs (Figure 10) More stable temperatures (Figure 11) Both controllers operate below thresholds (Figure 12) Results conservative b/c of low- powered CPUs and over- provisioned for cooling (i.e., large enclosure)