HETEROGENEOUS SENSOR NETWORK FOR DATACENTER WORKLOAD AND POWER MANAGEMENT JORGE ORTIZ CS ARCHITECTURES FOR INTERNET DATACENTERS OCTOBER 10, 2007 Sensing the Datacenter
Datacenters Consume Lots of Power Datacenter power consumption increasing Environmental Protection Agency (EPA) report shows power- consumption has doubled in last 5 years 1.5% of total U.S. Electricity Consumption in 2006 Projected to double again in next 5 years Datacenter under-provisioned for saving power Sensing and control separate from load balancer Protocols and applications are energy unaware What I propose: Couple the sensing and control with the load balancer/tasking decisions Dynamic Adjustment Graceful workload adjustment for saving power
Sensors Facilitate Dynamic Energy Accounting Include power-related input into the protocol and management loop Make use of equipment sensors already available to gather information about power consumption Power meters to attach to server/racks On-board temperature sensors In-band network monitors Wireless sensor network technology to include out-of- band monitoring infrastructures Single on-board sensors sometimes give wacky readings Array of sensors adds redundancy and improves accuracy Wireless motes ease deployment and data collection
Loose Ends Datacenter-scale workloads unavailable Monitoring machine room activity not at same scale as internet datacenter, but it’s a good start Power metering equipment needed A couple of one-outlet monitors already in use Access to 420A (Soda Hall Machine room) or the RadLab Machine room Direct access to specific machines in (either) machine room
Semester Plan Provision the Soda Hall machine room (420A) or the RadLab machine room with wireless sensors (temperature, humidity, etc.) Attach power meters to a set of servers in the machine room Setup process (top), network (netstat), and disk monitors (iostat) to determine distribution of machine activity Analyze gathered data Formulate a model that relates temperature and power consumption Analyze the relationship between component utilization and ambient temperature Machine-learning techniques for formulation of predictive models for graceful workload degradation