Hyperscale Computing, Fabrics and the Datacenter of Tomorrow
Total Page:16
File Type:pdf, Size:1020Kb
® Hyperscale Computing, Fabrics and the Datacenter of Tomorrow Larry Wikelius <[email protected]> Jon Masters <[email protected]> ® 1 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 How Does the Industry Define Hyperscale Today? q Webopedia.com: Hyperscale computing refers to the infrastructure and provisioning needed in distributed computing environments for effectively scaling from several servers to thousands of servers. q Whatis.techtarget.com: Hyperscale computing is a distributed computing environment in which the volume of data and the demand for certain types of workload can increase exponentially yet still be accommodated quickly in a cost-effective manner. q Wikipedia: Did you mean: Hyperspace? q Wordnik.com: Sorry – no definition found ® 2 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 http://www.hyperscale.com/ ® 3 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Overview l Hardware Technology Trends Leading the Way l System-on-Chip l Fabric Technology l Software Defined Networking l Software Technology Trends that Enable and Drive l I/O Driven Workloads l Big Data l Impact upon the Changing Datacenter l Where are we headed… ® 4 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – Single Core Performance We are reaching single-core performance limits -> 52% historical growth slows to 22% <- th ® Source: Computer Architecture 5 edition 5 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Historical Computing Efficiency Efficiency doubled every 1.57 years from 1946 - 2009 Source: Implications of Historical Trends in the Electrical Efficiency of Computing, 2011 ® 6 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – Mobile Technology q Mobile processors designed for energy efficiency q Target good performance/energy use trade-off q Pioneered System-on-Chip (SOC) integraon technology q Deep on-chip power/clock/frequency gang q Modern mobile devices also require performance q Tablets replacing laptops and desktops q MulA-core cellphones, with high end games q Mixture of asynchronous compute w/offload q Energy aware scheduling/power decisions ® 7 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 ® 8 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – System-on-Chip (SOC) q Originally an embedded systems concept q Allows very high density soluAons (e.g. cell phones) q Integrate customized value q Standard compute (ARM core) with customized devices q Specialized peripherals, offload engines, management q Package-on-Package allows stacking q Stack RAM/flash on top of the SoC q EnAre system in a few inches ® 9 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – System-on-Chip (SOC) Calxeda EnergyCore™ EnergyCore™ SOC Processor Complex Management Processor Complex Multi-core ARM® processors Engine integrated with high Advanced system, power bandwidth memory EnergyCore ARM controllers and fabric optimization Management and management for ARM Engine energy-proportional ARM Memory computing Controllers L2 Cache L2Cache ARM EnergyCore™ Fabric Switch I/O Controllers I/O Controllers Standard drivers, SATA, PCIe, EnergyCore Integrated high-performance standard interfaces. No Ethernet, SD/ Fabric Switch fabric converges internode surprises. eMMC communication, I/O, and storage networking ® 10 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 System Benefits of True SOC Integration Reduced Design Complexity 4 GB DRAM ECC mini-DIMMS Quad-core servers 4 SATA / Node Power, SATA, & Fabric Approximately 10” q A cluster of 4 servers q 20 was q No cables Image: Calxeda EnergyCard powered by Calxeda EnergyCore (source: Calxeda.com) ® 11 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – Integrated Management q Future datacenters will have millions of server nodes q None of these will have displays or serial cables q None of these will use exisAng “KVM” interface q Each SoC can integrate management features q Out-of-band IPMI, secure provisioning, etc. q Value-add opportuniAes for OEMs/vendors q Provisioning will eventually move to an injecAon-model q No thundering herd of servers hing PXE/DHCP q Inject (push) the OS/system images ® 12 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Fabric connectivity ® 13 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – Fabrics q Transparent, reconfigurable interconnect between many SoCs q No physical network ports to plug between systems q Redundant paths and rouAng opAmizaons built-in q Cluster level, server level and/or policy level controls q Virtualized IO and networking to all nodes q Replaces “top of rack” switch/cabling mess q Obviates need for one-to-one storage q Lower power and higher density q Enables true Soaware Defined Networks ® 14 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Fabric Topologies q Flexibility for market specific topology optimizations q North/south traffic (think web server) q East/west traffic (Fabric SANs, computational analytics) q Bisection bandwidth (classic HPC) q Appliances – Data flow optimized topologies q Different topologies spanning chassis Dual Fat Tree Butterfly Fat “Classic” Fat CLOS Network Tree 2D Torus Tree ® 15 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – Software Defined Networks (SDN) q Decouple data and control panes q Move control into (flexible) soaware q UAlize commodity general purpose compute q Programmable network layers q Dynamic reconfiguraon q Service innovaon (service-driven networks) q Virtualized infrastructure q OpenDaylight Project q Linux Foundaon SDN project ® 16 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – I/O Driven Workloads q Many modern workloads are dominated by I/O q Web Tier obvious example, but also data analyAcs q Storage and networking requirements escalang q Reduced requirement for big iron horsepower per-node q Performance in aggregate across sea of nodes q Connected through massive bandwidth fabrics q Focus on rack level throughput q Industry Benchmark example – Data Intensive Focus q Graph500 - hTp://www.graph500.org/ ® 17 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – Big Data q Distributed applicaon level storage becoming criAcal q Ceph/Gluster and other Open Source soluAons q Never throw any data away q MapReduce approach to processing huge data sets q “Map” large problems into smaller data chunks q “Reduce” results from many nodes into end result q Apache Hadoop is the well-known implementaon ® 18 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Technology Trends – IAAS - OpenStack q Hyperscale provisioning requires novel technology q Integrated management of whole data center q Nodes frequently re-installed (workload dependent) q “Physicalizaon” parAally displaces virtualizaon q On-chip management leverage for thin provisioning q OpenStack becomes the “Datacenter OS” of the future ® 19 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Changing Data Center – Energy Efficiency q Data center energy costs are rising q Power Usage EffecAveness (PUE – lower numbers beTer) q Performance-per-Wa as a key compute measure q Hard limits on per-rack power consumpAon q Prevent business as usual approach q Workarounds q Infrastructure equipment overhead (networking, etc.) q Also associated HVAC infrastructure q Carbon dioxide and climate concerns ® 20 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Changing Datacenter – Hyperscale q Rethink convenAonal datacenter blade approach q Integraon (SoC, Fabric, etc.) q Dramacally reduce energy consumpAon q 1000+ nodes possible in a rack today q 10’s of thousands within 3-5 years q Performance in aggregate, not per core q Many simple nodes replace single points of failure q Hyperscale designs leverage simpler hardware ® 21 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Changing Datacenter – Vertical Integration q Emerging shia toward purpose-built verAcal designs q Leverage standard (ARM) compute cores q Integrate I/O and peripheral devices on SoC q Integrate applicaon specific accelerators (offload) q Open Source principals applied to hardware design q Facebook OpenCompute provides reference designs q Leverage high-value ODMs ® 22 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Changing Datacenter – Fabrics q The “real story” behind ongoing datacenter revoluAon q Next big baleground in server design q Fabric Wars? q Disaggregaon of resources q Separate Compute from RAM, and Storage q UAlize dynamically configured million node fabrics q OpenCompute blazing a trail ® 23 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Where are we Headed… q Million node systems become a reality q New opportuniAes in Compute/Big Data/AnalyAcs q DisrupAve impact across the Data Center q Future datacenters are dark places nobody goes q Flood fill with SoC/Fabric designs q Failure-in-Place up to set threshold q 3-5 year disaggregated depreciaon ® 24 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 Closing – Now How do we Define Hyperscale? q Fine grain complete building blocks q “Infinite” HW resource – at a scale not previously seen q Solution Optimized q Connectivity and throughput drives implementation q Breakthrough in overall performance and performance efficiency across Data Center ® 25 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 The Real Future for Energy Efficiency Computing Image: An HP Redstone 32-bit ARM server under JCM pedal power at RH Summit ® 26 Hyperscale, Fabrics, and the Datacenter of Tomorrow | End User Summit 2013 .