WIXLIB

Percy Tzelnic – SVP, EMC FellowSorin Faibish – DE, ArchitectFast Data GroupOffice of the CTOEMCEMC CONFIDENTIAL—INTERNAL USE ONLY1

FlashDynamically Loadable NS (DLN)2 TIERS Roots In Exascale ResearchFast TierJob2 TIERSTM EMC collaboration with DoE andIndustry Consortia in FastForwardExascale–Fast acceleration tier Performance of flash–Large capacity tier Retention and capacity of object store–Global POSIX namespace over onetrillion objectsObject StoreDLNIndexTablePackedDLNs2T metadata2T file dataCapacity TierEMC CONFIDENTIAL—INTERNAL USE ONLY2

Fundamental Change In Storage Architecture Contemporary Storage Architecture is being disrupted: Flash replaces disk for 100x performance (Flash Array) Cloud replaces disk for 100x capacity (Object Store) Capacity disks from arrays move to the cloud, leaving a Flash only Fast Tier onpremise, and an Object Store only Capacity Tier, in the Cloud or sometimes onpremise We can no longer package Performance and Capacity in one box at anattractive price/value point Split the two, hence 2 TIERSTM (Fast and Capacity Tier)EMC CONFIDENTIAL—INTERNAL USE ONLY3

Instantiation Of 2 TIERS In EMC In EMC, Fast Tier can be instantiated in the network byDSSD, or in the servers by ScaleIO local flash 2 TIERS is Software Defined Storage with two keycomponents: PFS – Parallel File System (OrangeFS) and Syncer (built by EMC for the DOE FastForward program,for Exascale I/O stacks)“The DSSD D5 Storage Appliance thisbrings me to some future technologies thatEMC was showing in their Innovation exhibit.There they were showing what they calledTwo-Tiers model .”“The Two Tiers model provided consistentsubsets of both data and metadata betweenthe hot edge and cold core storage capacity(hence two tiers). The company said it wasseeing performance and cost benefits fromthis solution prototype for IO and capacityintensive applications.”Tom Coughlin, Intersect /08/emc-cloud-storage-flash-memory-and-beyond/ In EMC, the Capacity Tier can be instantiated by ECS or Isilon EMC 2 TIERS software runs in the Fast Tier, presents the POSIX API andNamespace to apps and maps the apps data into objects on the Capacity Tier, withpolicy driven tiering between the twoEMC CONFIDENTIAL—INTERNAL USE ONLY4

2 TIERS Was Designed For The 3rd PlatformBut It Provides Equal Support To 2nd Platform Apps!Disaggregate the monolithic memory / storage / IO Stack and recast itinto loosely coupled “Fast Tier” and “Capacity Tier”, to enable Independent Scaling:–Scale-out for Fast Tier, O(100 to 1,000)–Hyperscale for Capacity Tier, O(100,000) POSIX API and Namespace–Required by the majority of 2nd platform apps (at a scale-out level lower than for a3rd platform infrastructure)EMC CONFIDENTIAL—INTERNAL USE ONLY5

How Does 2 TIERS Work?Read-only, Read-through TranslationService on Local FUSE File SystemApp Local Storehyperstub2T metadatapromotedmodifiedTIME T1:Persist DLN d,version v 1TIME T0:Promote DLN d, version v2T file dataCapacity Tier at Time T0EMC CONFIDENTIAL—INTERNAL USE ONLYFlashnewTIME T0:Load DLN d, version vITPackedDLNsFast Tier on Distributed OrangeFSCapacity Tier at Time T16

Tiering Data And MetadataMetadata01DLN Index TableOID 0x0114 a0fd27a480DLN bucketEMC CONFIDENTIAL—INTERNAL USE ONLYData28 0a0ff0cb3c44 67 4210 90 890a f0 3c3df17

How Does Metadata Tiering Work?Similar To A UNIX FFS File System Expose a global namespace view to the app, similar to a File System Use a pre-defined Global Unique OID for the DLN Index Table(Cassandra KVS); similar to a SuperBlock of a File System Each Entry in the Index Table points to a DLN; similar to inodes fordirectories in a File System Each DLN points to a partial view of global namespace; similar to asubtree of a File System Each namespace view has a pointer to an object; similar to file inodesin a directory in a File SystemEMC CONFIDENTIAL—INTERNAL USE ONLY8

EMC 2 TIERS – One Of Many Similar Approaches Since June 2014, while EMC develops 2 TIERSTM, an ever increasing number ofsimilar projects have emerged in the industry, both in EMC and outside(university research, new ventures, etc.)––––––––––MarFS (LANL) – the most similar to 2 TIERS , in production for Campaign VFSBetrFSGiraffaFS This is good confirmation of two widely resonating concepts:––Object Storage for Capacity, Flash for Performance!Users think in folders, not objects! . They need a File System Namespace!EMC CONFIDENTIAL—INTERNAL USE ONLY9

Differentiation Of EMC 2 TIERS Unique Characteristics Of 2 Tiers 1. Single Global Namespace with Dynamically Loadable Namespaces (DLNs)2. Tiering of both Data and Metadata3. Fast Tier Performance Target: greater than 10X Capacity Tier4. Direct access (read-only) to the Capacity Tier, bypassing the Fast Tier5. 2 TIERSTM provides Tiering and Non-Tiering modes6. No client changes required7. No changes to the EMC products required for EMC instantiationEMC CONFIDENTIAL—INTERNAL USE ONLY10

2 TIERSTM: Local Or Network Fast Tier ExamplesLocal Fast TierNetwork Fast TierApp Cluster(Compute Servers) 2T Servers Local FlashApp Cluster(Compute TiersApp2Tiers2TiersRDMAFlashDSSDIsilon, ECSAppCapacity Tier2T ServersDSSDIsilon, ECSCapacity TierNote: Compute Server interconnect should be RDMA, for best performanceEMC CONFIDENTIAL—INTERNAL USE ONLY11

A Possible EMC Product PackagingFast Tier2Tiers, in8 Servers4 BladeGbE4 Blade4 Blade4 Blade10 GbE10 GbE4 Blade4 Blade10 GbE 48p10 GbE 48p4 Blade4 Blade1 GbE 48p4 Blade4 Blade4 Blade60 Disk4 Blade4 Blade4 BladeFast TierDSSDComputeNodes (44)4 Blade60 Disk4 BladePCIeDSSDFlash Array60 DiskBrace60 DiskService TrayPDU4 Blade4 BladeECSCapacityTierIsilonCapacityTier60 Disk4 Blade4 Blade10 GbE 48p60 Disk10 GbE 48p1 GbE 48p4 Blade4 Blade4 Blade60 Disk4 Blade4 Blade4 Blade4 Blade60 Disk4 BladeBraceNot UsedECS U4000ECI Caspian withDSSD & IONs (EMC-IOD) (3.9PB)EMC CONFIDENTIAL—INTERNAL USE ONLY12

OrangeFS: EMC Choice For 2 TIERSTM Stateless design of underlyingPVFS2 Light weight Linux kernel module,multi-threaded client Performance comparable to other PFS HDFS with JNI client; support forWindows, Mac Modular design Abstract key-value interface formetadata Abstract storage interface for data Abstract networking allows RDMA, IP Client changes NOT requiredEMC CONFIDENTIAL—INTERNAL USE ONLY Future roadmap: OFS V3 –changes for CloudyCluster (Cloud PaaS); already deployedon AWS OrangeFS is maintained anddeveloped by Omnibond,Clemson, SC Agile and responsive open sourcecommunity Committed to open sourcecommunity development History of 4-5 years in production, atmajor customers13

EMC 2 TIERSTM On Omnibond CloudyCluster On AWS 2 TIERSTM CustomerDemo built onCloudyCluster to hostPOSIX apps as POC onAWS Several LargeCustomers of EMC haveasked for access waitingto 2 TIERSTM on AWS After for POC ready,significant dev workremains Initial Custom AMI builtby EMC with 2 TIERSTMand CloudyClusterEMC CONFIDENTIAL—INTERNAL USE ONLYHPC Compute GroupsDynamoDBSchedulerHighly Available 2 TIERSTM OrangeFS StorageOrangeFSWebDAV2 TIERSTMManagementInstanceS314

Self Service Elastic HPCCCQHPC JobCreate a fully operational HPC Cluster inminutes, complete with: Storage: OrangeFS on EBS, S3, EFS Compute: Job Driven Elastic Compute through CCQ Scheduler: Initially Torque with CCQ MetaSchedulerHPC Libraries:Boost, Cuda Toolkit, Docker, FFTW,FLTK, GCC, Gengetopt, GRIB2,GSL, Hadoop, HDF5, ImageMagick,JasPer, NetCDF, NumPy, Octave,OpenCV, OpenMPI, PROJ, R, Rmpi,SciPy, SWIG, WGRIB, UDUNITSHPC Software:Ambertools, ANN, ATLAS, BLAS, Blast,Blender, Burrows-Wheeler Aligner,CESM, GROMACS, LAMMPS, NCAR,NCL, NCO, nwchem, OpenFoam, papi,paraview, Quantum Espresso,SAMtools, angeFS HPC Parallel StorageDDBEFSS3 All from an easy to use Web UI from mobile, tablet or desktopEMC CONFIDENTIAL—INTERNAL USE ONLYAvailable now in the15

Accessing 2 TIERS On AWSView available2 Tiers clustersEMC CONFIDENTIAL—INTERNAL USE ONLY16

List Of 2 TIERS ClustersSingle instance2 Tiers clusterOn AWS WestEMC CONFIDENTIAL—INTERNAL USE ONLY17

View Of 2 TIERS Cluster On AWSControl NodesCompute NodesOrangeFS NodesManagement NodesEMC CONFIDENTIAL—INTERNAL USE ONLY18

EMC CONFIDENTIAL—INTERNAL USE ONLY19

EMC CONFIDENTIAL—INTERNAL USE ONLY20

EMC CONFIDENTIAL—INTERNAL USE ONLY21

EMC CONFIDENTIAL—INTERNAL USE ONLY22

In EMC, the Capacity Tier can be instantiated by ECS or Isilon EMC 2 TIERS software runs in the Fast Tier, presents the POSIX API and Namespace to apps and maps the apps data into objects on the Capacity Tier, with policy driven tiering between the two Instantiation Of 2 TIERS In EMC