Dave Bradley Rick Harper Steve Hunter 4/28/2003 CoolRunnings

Overview  Joint effort between Research and xSeries Development  Predictive algorithm that measures and predicts workload and determines when to place servers in a low power state  Objective is to minimize energy consumption, unmet demand, and power cycles Automatically adapts to short term and seasonal workload variations Automatically adapts algorithm "gains" to workload dynamics Energy savings of 20% or more can be achieved  Demo developed to illustrate benefits  Findings to be published in IBM Journal of Research and Development later this year

Problem  Power consumption and dissipation constraints are jeopardizing the ability of the IT industry to support the business demands of present-day workloads It can be difficult to get power into and heat out of large systems Power shortages such as recently experienced in California impose further, unpredictable, constraints Server complex size is growing dramatically in response to electronic commerce and web hosting computational needs Power consumption is growing to well over 100 watts per processor

Observations  Electronic commerce and web serving workloads have certain characteristics that make them amenable to predictive system management techniques that can effectively reduce power consumption They usually support periodic or otherwise variable workloads, with the peak workload being substantially higher than the minimum or average workload The peak workload has been seen as ten times the minimum (and sometimes average) workload

Observations  These workloads are highly parallel and easy to load balance A typical web serving system has a large number of web servers fronted by a load-balancing “IP sprayer”, which provides a single IP address to the outside world, and dispatches requests from the outside world to the many web servers in the complex to balance the load among them The IP sprayer sends a given request to the server having the lowest utilization, and, in turn, the servers keep the IP sprayer updated with their utilization, response time, or other indication Workload is routed around failed servers, and the users having transactions or sessions on a failed server can click again and their request will go to another server.

IP Load Balancing Application Traditional Load Balancing Model Load Balancer Internet Step 1: Document request (IP-SVA) Step 2: Web server selection - Appropriate IP Address inserted Step 3: Packet forwarding Step 4: Document response (IP-SVA) - Response passes back through Load Balancer Network Switch Server 2 Server N Server 1 4 Client 2 4 Characteristics  Efficient use of Server Resources  High Availability/Scalability  Bandwidth Limited on Return Path (i.e., Response Traffic ) through the Load Balancer Single IP Address

The CoolRunnings Algorithm  CoolRunnings exploits this environment to manage power based on measured and predicted workload, such that both unmet demand and power consumption are minimized. It does this by: Measuring and characterizing the workload on all the servers in a defined group Determining whether any servers need to be powered on or off in the near future, by assessing the current capacity relative to the predicted capacity needs, Manipulating the existing system and workload management functions to remove load from servers to be turned off, and Physically turning on or off (or transitioning into and out of standby or hibernation) designated servers using existing system management interfaces

Application of CoolRunnings Power Management in IP Spraying Schema

Application of CoolRunnings Up to 14 Processor Blades  Modular, Scalable Density with Performance  7U Mechanical Chassis Integrated Network Infrastructure  Switching with point-to-point blade connections Affordable Availability  Redundant, Hot-swappable blades and modules Advanced Systems Management  Integrated service processor Power Management with BladeCenter

Application of CoolRunnings Power Management with Virtual Machines

Demo

Summary  CoolRunnings is a real-time power management algorithm that is applicable to a complex of servers having a parallelizable and migratable workload  The algorithm is designed to minimize power utilization, unmet demand, and server power cycles  Experiments with the CoolRunnings algorithm on /database and web serving workloads show that energy savings of 20% or more can be readily achieved, with typically less than 1% of unmet or deferred demand and an average of one power cycle per server-day.  If the workload has a high dynamic range, or if the system is substantially overprovisioned, much higher energy savings are obtained.