Meikel Poess Oracle Corporation
Analytical Power Consumption Model Based on nameplate power consumption Nameplate power is conservative estimate Model adjusts nameplate to yield realistic numbers Estimates peak power consumption during steady state workload Developed for OLTP and Decision Support workloads Validated with measured power numbers of TPC-C, TPC-E and TPC-H benchmark results Power estimates are very close to measured numbers Performed long term study on 64 x86 based systems Meikel Poess2
Energy Consumption Of TPC-C Systems Slope of x increase over 7 years 40% lower than performance increase in the same period 3Meikel Poess Slope of x increase over 7 years Power consumption is increasing
Power Performance Analysis x-axis: performance [Q/h] y-axis: Total Power [KWh] 6 different configurations 6 disks 14 disks 32 disks 100 disks SSDs In-Memory Shows energy efficiency of systems Meikel Poess4 Power-Performance Quadrant
Impact of Performance Techniques To Power-Performance 5Meikel Poess Shows impact of compression to Power- Performance 5 different configurations 6 disks 14 disks 32 disks 100 disks SSDs Each configuration defines a vector Vector indicates how much performance was gained how much power was saved Other techniques CPU speed adjustment Power-Performance Quadrant
Research Questions Data placement What data to store on fast vs. slow disks or on SSDs or when to keep it in Memory? Can data be placed intelligently depending on a dynamic workload Power Management If we can improve system performance at a higher rate than we need to increase power consumption, what do we do with the idle resources? Service level agreements How can we incorporate power aspects into service level agreements? 6Meikel Poess