ACM Symposium on Cloud Computing (SoCC 2015), August 27–29, 2015, Kohala Coast, HI, USA CoolProvision: Underprovisioning Datacenter Cooling Overview Design Results Architecture Datacenter Cooling Traditional cooling accounts for 40% of construction costs High energy costs (location-dependent) Modular design Inputs 1. High-level thermal models 2. Historical weather (temp, RH%) 3. Expected workload profile 4. Server power models 5. Power – performance trade-offs Outputs 1. Cheapest cooling technology 2. Cheapest cooling size Constraints Upper inlet temperature and RH% Maximum workload degradation Ioannis Manousakis, Thu D. Nguyen {ioannis.manousakis, tdnguyen}@cs.rutgers.edu Sriram Sankar [email protected] Epoch – based optimization Tunable horizon Method: SQP Optimization Design goals: Systematic way to reduce datacenter cooling within perf. constraints Trades capital costs, energy costs, hardware reliability and performance Four Representative Locations Workload Management Policies Operational Example Inigo Goiri and Ricardo Bianchini {goiri, ricardob}@microsoft.com Reduce server power and heat during harsh conditions 1. High outside temperature 2. High relative humidity 3. High load Generic workloads DVFS Interactive and VMs Consolidation Deferrable analytics Job deferring a) Traditional Cooling (Chillers) b) Direct Evaporative Cooling Implementation in MATLAB (Simulator) Trained models with real small factor A/C and evaporative cooler (Parasol) Benefits: Reduced capital costs and drastically improved NVPs Workload- and location-aware cooling deployments Inlet Temperature ( o C) Austin TX - DX A7C 080 Newark NJ- DX A7C Singapore - DX A7C [1] Full-Load 1 [2] Full-Load 2 [3] Batch [4] Batch - Defer [5] Interactive 1 [6] Interactive 2 Austin TX - Evap 28 Newark NJ-Evap Singapore - Evap 28 30 35 28 30 35 37 30 35 28 30 35 28 30 35 28 30 35 28 30 35 085 180