Choosing Right All-Flash-Array Aleksandr Shvadtshenko Sr. Systems Engineer aleksandr.shvadtshenko@emc.com 1
INDUSTRY TRENDS A FEDERATION PERSPECTIVE HOW 5 TRENDS ARE INTERCONNECTED Next Generation Applications Trend PaaS, Mobile, Hadoop Which Will Enable Hybrid Clouds Which Will Be Used As Building Blocks Which Will Be Deployed In CI Hybrid Cloud Trend Software-Defined Data Center Trend Converged Infrastructure Trend Elastic, Agile, Data Center Running Next Generation Apps 2 Key Changes To The Underlying Storage Market Flash Trend New Levels of Performance COTS Trend New Levels of Efficiency 2
Annual Bookings ($M) $1BN 900 800 700 600 500 400 300 200 100 0 0 1 2 3 4 5 6 7 Years Since Product Availability 3
Evolution of All-Flash Array Architectures This research shows that flash will become the lowest cost media for almost all storage from 2016 and beyond, and that a shared data philosophy is required to maximize the potential from both storage cost and application functionality perspectives. http://wikibon.org/wiki/v/evolution_of_all-flash_array_architectures 4
The All Silicone Data Center 5
Why Flash? Challenge #1: Latency Bad user experience Application timeout Lost time = money 6
Why Flash? Latency cause: CPU HDD performance gap CPU IS FAST AND GETS FASTER DISK IS SLOW AND DOES NOT GET ANY FASTER LATENCY 2000 2010 2020 100x 800x 10000x 7
Why Flash? Solution: Flash will close the performance gap CPU IS FAST AND GETS FASTER FLASH IS FAST AND GETS FASTER 2000 2010 2020 100x 800x 10000x 8
XtremIO Sudden Impact - SQL Server 9
Gen 1 Flash Systems Legacy Disk Arrays The Evolution of FLASH Arrays Scale Out All Flash Arrays Multi Controller (16-XtremIO) Flash Optimized Hybrids Dual Controller Limited Scale Single Workloads Some Data Services Scale Up All Flash Arrays Dual Controller Limited Scale Single Workloads Sometimes Data Services Compromise Between Performance & Efficiency - Data Svcs Petabyte Scale Mixed Workloads In-Line Data Services Lowest TCO Highest Performance All Flash Data Center Gen 1 Gen 2 Gen 3 Gen 4 Source: Wikibon December 2014 10
Changing Tides Gartner 2013 AFA Market Share IDC 1H 2014 AFA Market Share Revenue ($M) Share (%) EMC 112.3 22.6 Pure 90.0 18.3 IBM 82.9 16.7 IDC 1H 2014 HFA Market Share Revenue ($M) Share (%) EMC 1,575.8 35.5 NetApp 891.8 20.1 Hitachi 521.2 11.7 11
Why Architecture Matters: Key Ingredients Software-Defined SCALE-OUT Inline & Unstoppable DATA SERVICES Thin Provisioning EMC Portfolio INTEGRATION sub 1ms latency Deduplication Compression Flash Data Protection Encryption Writeable Snapshots VPLEX SRM PowerPath RecoverPoint Linear Scale IOPS & Capacity Data Reduction Efficiencies HA/DR, Management, Converged CONSISTENT & PREDICTABLE PERFORMANCE @ SCALE (NO SYSTEM-LEVEL GARBAGE COLLECTION) 12
Avoiding Bad Design Choices 5 Common Things AFAs Borrow from Disk Log Structuring Uneven SSD wear Post-Process Performance inconsistency RAID Par ity Par ity Par ity Par ity SSD Wear - Write Amplification Metadata De-stage Performance inconsistency Asymmetric Controllers A P Uneven Resource Utilization XTREMIO DOES NOT USE ANY OF THESE 13
XTREMIO DATA PROTECTION Designed for SSD No legacy RAID baggage Highly efficient- only 8% overhead Lowest write amplification SSDs may fail in place No configuration No hot spares Adapts to failures High Performance RAID 1 Good Capacity Utilization RAID 5 XDP Superior Protection RAID 6 Copyright 2014 EMC Corporation. All rights reserved. 14
Typical AFA System Level Garbage Collection Array Controller Huge back-end I/O amplification initiated by array controller Cannot defer process when array must free up space Big tax on array controllers to manage limits capacity 15
XtremIO No System Level Tax XtremIO Controller No Garbage Collection ASIC ASIC ASIC ASIC ASIC ASIC ASIC ASIC ASIC Performed by Each SSD Controller ASIC SSD controllers have ideal knowledge of the NAND Zero back-end I/O initiated by the array controllers No tax on array controllers 16
ACTIVE ACTIVE 16 ACTIVE CONTROLLERS 150K SCALE IOPS SCALE 2M IOPS Scale Up XtremIO Scale Out SAN Controller 14 Controller 13 Controller 16 Controller 15 Controller 12 Passive Active Controller 10 Controller 9 Controller 11 Flash Flash Flash Controller 6 Controller 5 Controller 8 Controller 7 Controller 4 Controller 2 Controller 3 Controller 1 17
Why Customers Choose XtremIO Latency Latency Consistent High Performance & Low Latency Actual customer data Workload: OLTP with 8KB Block Size Latency Spikes 2-4.5ms = Gen 3 Flash Various latency spikes <1ms = XtremIO Consistent & Predictable (~0.5ms) System-Level Garbage Collection Impact Worse than Disk Latency 20-40ms = Gen 3 Flash Latency increases with capacity <1ms = XtremIO Always Consistent & Predictable XtremIO X-Brick Traditional Flash Array 18
XtremIO X-Brick Cluster Building Block Active Controller 1 Active Controller 2 32 CPU CORES 256GB RAM Infiniband RDMA 256GB RAM 512 GB RAM 2 x FC 2 x iscsi SAS 2.0 SAS 2.0 4 HOST PORTS 2 x FC 2 x iscsi 25 emlc SSDs 19
How Does It Scale? X-Brick 1 SHARED MEMORY METADATA Unique User Data SHARED MEMORY METADATA SHARED MEMORY METADATA Unique User Data SHARED MEMORY METADATA X-Brick 3 RDMA FABRIC SHARED MEMORY METADATA Unique User Data SHARED MEMORY METADATA X-Brick 4 SHARED MEMORY METADATA Unique User Data SHARED MEMORY METADATA X-Brick 2 20
Product Family, On-Demand Linear Scale-Out 150K mixed IOPS 250K read IOPS <1ms Latency From 2 16 N-way Active Controllers 1.2M mixed IOPS 2M read IOPS <1ms Latenc PBs Capacity Starter 5TB 5 320TB physical, PBs effective capacity based on 10, 20, & 40 TB X-Bricks SCALE-OUT <1ms LATENCY RICH DATA SERVICES NO TUNING 21
Agile Writeable Snapshots 100% IN-MEMORY Any topology Instant creation Instant deletion 100% SPACE EFFICIENT No space reservations No metadata bloat 100% PERFORMANCE Identical read IOPS Identical write IOPS Identical latency 100% OPTIMIZED Identical data services Always on, always inline INCREDIBLE SCALE Instant application clones to petabyte scale UNMATCHED Use XtremIO where allflash arrays were never before viable 22
Today s Applications Environment JUST LIMITED TEST/DEV COPIES Brute Force Copy PRODUCTION Brute Force Copy DATAMART COPIES ONE APPLICATION 6 DATABASE COPIES 3 ARRAYS/POOLS 1 USE CASE FOR FLASH 23
Game-Changing Consolidation, Agility SCALE-OUT IOPS IN ABUNDANCE FREE, FAST DEV/TEST FREE, FAST ANALYTICS XTREMIO FLASH FOR ENTIRE APPLICATION DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 DEV INST 1 FIN COPY FIN COPY REPORTS MORE BUSINESS PRODUCTIVITY DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 DEV INST 2 OPS COPY OPS COPY ANALYTICS DEV INST 3 DEV INST 3 DEV INST 3 DEV INST 3 DEV INST 3 DEV INST 3 x TEST/DEV COPIES DEV INST 3 DEV INST 3 DEV INST 3 DEV INST 3 1 0 DATABASE INSTANCES HIGH PERFORMANCE PRODUCTION 1 SALES COPY XTREMIO CLUSTER SALES COPY DATAMART COPIES 0 TRENDS FASTER APPLICATION DEVELOPMENT TIMES BRUTE FORCE COPIES 24
XtremIO Simplicity & Automation 1CREATE VOLUMES 2CREATE INITIATOR GROUPS3MAP VOLUMES Management Impact: Zero Planning & Tuning No storage skills No certifications Provision in seconds! Integration: VMware vcenter EMC ViPR, ESA, ESI, SRM Microsoft Hyper-V App Consoles: Oracle, SAP, System Center, etc REST API & CLI 25
XtremIO: Leading Use Cases Database & Business Apps Consolidate production, dev/test, BI/analytics/reporting instances Solve the toughest SLA challenges TCO savings across storage, servers, app licensing Virtual Desktops & DaaS Uncompromising user experience, at scale and all desktop types Scale-out as VDI grows for any mix of desktop types Lowest Opex & Capex, <$100/Desktop Cloud & Virtual Servers Infrastructure Private cloud mixed workload consolidation Mission-critical app virtualization Application-as-a-Service Software Dev/Ops Infrastructure Cloud-scale TCO savings for Capex and Opex Consistent Performance Best User Experience No Complex Setup/tuning Provisioning Simplicity EMC Portfolio Synergy 26
Data Reduction Guidelines Below are a few examples for data reduction rates: Use Case Dedupe Ratio Compression Ratio Data Reduction Ratio VDI (full clones) 6:1 10:1 1.3:1-1.5:1 7.8:1 15:1 VDI (linked clones) 1.5:1 1.3:1-1.5:1 2:1 2.3:1 Virtual Servers 1.5:1 3:1 1.3:1 2:1 2:1 6:1 SQL Server 1.1:1 1.5:1 1.8:1 1.6:1 2:1 Oracle DB 1.1:1 2.2:1 2.5:1 2.4:1 2.8:1 * The rates listed above are only examples and do not necessarily reflect real life values 27
XPECT MORE PROGRAM * EMC IS REDEFINING THE STORAGE LIFECYCLE MAINTENANCE PRICE PROTECTION 3-YEAR MONEY-BACK WARRANTY FLASH ENDURANCE PROTECTION 7 YEARS 2014-2017 7 YEARS * For qualifying customers through December 31, 2015. See EMC.com/XpectMore for details, terms and conditions. 28
The Rise of Server SAN http://wikibon.org/wiki/v/the_rise_of_server_san 29
2ND AND 3RD PLATFORM ARCHITECTURE Blades, Virtualization, Arrays COTS, Storage Software 30
EMC SCALEIO SOFTWARE-DEFINED, SCALE-OUT Utilize commodity hardware for block storage Run storage and applications on the same servers Experience flexible & scalable performance & capacity on demand Choose your operating system, hypervisor and media 31
ScaleIO Hyper-converged architecture Flash SSD HDD RAID Cache RAM Bare Metal KVM VMware Hyper-V Compute Network ETH/IB Storage 32
ScaleIO Hyper-converged architecture Flash SSD HDD RAID Cache RAM Bare Metal KVM VMware Hyper-V S C S C S C S C S C S C S C S C S C S C 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 225K IOPS 10 TB 2,250,000 IOPS 100 TB 33
agnostic HDDs PCIe Flash Bare Metal SSDs 34
ScaleIO architecture: Mapping storage: RAW DEVICE or partition 35
ScaleIO architecture: Mapping storage: RAW DEVICE or partition 36
ScaleIO architecture: Mapping device: FILE WITH PREPARE_FILE utility Utility Distributed with SDS Windows Software Prepares a File to use in ScaleIO..\create_file_storage.exe --create_file --size_gb 100 --file_name c:\scaleio1.bin Can be used in the CSV Deployment Manager 37
ScaleIO architecture: scaleio hyperconvergance: resource utilization Impact on server is below 10% CPU load Usually far below Maximum IOPS load on SSD pool Maximum IOPS load on 10K pool 38
ScaleIO Automatic data rebalancing Dynamically add, move, remove storage and compute resources on the fly with no downtime Automatic volume rebuilds and rebalancing Mix Server brands Configurations OS platforms (physical & virtual) Devices (SSD, HDD, PCIe, LUNs, partitions, files) No capacity planning required! 39
Elastic architecture Add, remove, re-allocate, on the fly BALANCED BALANCED BALANCED Auto-balance of resources across nodes/clusters Auto-rebalance when resources are added Auto-rebuild when resources fail or removed BALANCED BALANCED Easier Capacity Planning No Migrations 40
ScaleIO Scalability Scale to thousands of nodes Add devices and servers to increase capacity and performance Storage growth always aligned with application needs 16PB per cluster 2048 nodes in one single cluster 64 devices per node 6TB device 41
ScaleIO PERFORMANCE Scales Linearly ~31M IOPs is about 8X better than the latest high-end HDS storage (G1000) for a fraction of the cost If we were to extrapolate the lines, we would get ~180M IOPs for 1024 Nodes (~46 G1000 systems ) 42
ScaleIO SUB-MILLISECOND LATENCY Measured from the application perspective Typical read latency figures with flash and 10GbE Write figures are ~1.5X to 2X 4KB: 304 usec 8KB: 344 usec 128KB: 778 usec 43
ScaleIO System wide visibility, cost effective management at scale Automated installation scripts and easy configuration Manage the entire data center stack from single UI VIEW AGGREGATE CAPACITY PERFORMANCE MB/S & IOPS Monitor HW/SW failure recovery no administrative intervention required Options: CLI, UI, REST, ViPR Controller, ViPR SRM, vsphere, OpenStack SCALEIO NODE COUNT READS/ WRITES 44
Enterprise Grade PERFORMANCE INTEROPERABILITY MONITORING FAULT TOLERANCE RESILIENCY Flash Cache & XtremCache OpenStack & vsphere SNMP & Call Home Rack Level High Availability RecoverPoint APP 1 APP 2 APP 3 MULTI-TENANCY PROTECTION SECURITY ELASTICITY EFFICIENCY Protection Domains & QoS Controls Writeable Snapshots Data Masking Automatic Rebalance & Rebuild Thin Provisioning 45
Use Cases ANY APPLICATION THAT USES BLOCK STORAGE TRADITIONAL NEXT GEN ANALYTICS APPLICATIONS 46
EMC ScaleIO node Bundled EMC commodity servers and ScaleIO software Fully architected Simplified planning and building Pre-validated, tested and configured Single vendor fully supported 47
SUPREME ELASTICITY UNPARALLELED FLEXIBILITY MASSIVE SCALABILITY EXTREME PERFORMANCE COMPELLING ECONOMICS ATTAIN 8x better performance than TRADITIONAL SAN Scales To 1k+ NODES EFFICIENCY 50 % Over Traditional SAN Up To 30-60 % TCO Savings 48
.Next Steps Visit EMC.com/ScaleIO where you ll have access to whitepapers, videos, demos etc. Download Free & Frictionless version of ScaleIO Join the ECN ScaleIO Product Community Follow Us on Twitter @EMCScaleIO 49
ScaleIO POC results in Estonia 50
2ND AND 3RD PLATFORM ARCHITECTURE Blades, Virtualization, Arrays COTS, Storage Software 51
ViPR CONTROLLER CLOAKS & MANAGES ALL DIAMOND GOLD SILVER BRONZE Control Abstraction VMAX VNX VBLOCK XtremIO Isilon EMC ScaleIO 52
ViPR service catalog 53
54