Collective Intelligence for Data Center Operations Management Xiaojun (XJ) Liu, Chief Scientist Spark Summit 2013 Outline ● Operations Management SaaS ● Experiences with Spark ● Ongoing Work Operations Management SaaS Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 3 of 11 Data Pipeline Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 4 of 11 The Data We Collect ● Configuration, performance, tasks and more ● From virtual machines, servers, networks and storage ● 100 billion+ metric samples per day ● On average 1.3 million properties per datacenter Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 5 of 11 Cross-User Analysis Setup ● ● ● Spark Cluster (EC2) Resource Sizing Consolidation Ratio Utilization … (Standalone App in Scala) Master Tesla CloudPhysics library for Spark Data query ● Data Repository (S3) References Analysis App: Worker Data History (HBase) range, firstAfter, … Extraction ● config/perf/task Segment ● Status Server (Play! App) by running vm or host count Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 6 of 11 Analyses We Have Done ● VM configured resources and utilization ● Server CPU and memory utilization distribution ● Storage array and interconnect adoption ● Virtualization product feature adoption Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 7 of 11 VM vCPU and Memory Size Distribution Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 8 of 11 Experiences with Spark on EC2 + Quick getting up to speed + Great EC2 support + Tolerating variations in task execution time - Variations in EC2 instance performance Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 9 of 11 Next Steps ● ● ● Create RESTful API for frequently used analyses to make update and consumption easier Utilize Shark and MLbase for cross-user data analysis Utilize Spark Streaming for near-realtime performance data analysis Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 10 of 11 Thank You! www.cloudphysics.com [email protected] @cloudphysics Xiaojun (XJ) Liu [email protected] Spark Summit Dec 2, 2013 Copyright © CloudPhysics 2013 Slide 11 of 11
© Copyright 2026 Paperzz