HPC Advisory Council September 2012, Malaga CHRIS WEEDEN SYSTEMS ENGINEER
WHO IS PANASAS? Panasas is a high performance storage vendor founded by Dr Garth Gibson Panasas delivers a fully supported, turnkey, hardware and software clustered storage solution with multiple protocol support Panasas offers a highly scalable, high performance parallel file system The Panasas solution is built upon the Object Storage Architecture standard Panasas is designed to release the true performance of compute clusters Panasas simplifies system management and reduces the cost of administration COMPANY CONFIDENTIAL 2
EVOLUTION OF NAS What is Parallel Storage? NFS Clustered NFS File Server File Server File Server Parallel NFS DAS: Direct Attached Storage NAS: Network Attached Storage Clustered Storage: Multiple NAS file servers managed as one Parallel Storage: File server not in data path. Performance bottleneck eliminated. COMPANY CONFIDENTIAL 3
MARKET SEGMENTS Aerospace Automotive Biosciences Energy Fluid Dynamics (CFD) Structural Mechanics Crash Analysis Fluid Dynamics Acoustic Analysis Genomic Sequencing Molecular Modeling Seismic Processing Reservoir Simulation Interpretation Finance Government Industrial Mfg University Credit Analysis Risk Analysis Portfolio Optimization Imaging & Search Weapons Simulation Weather Forecasting EDA Simulation Optical Correction Thermal Mechanics Materials Science Bio Sciences Weather COMPANY CONFIDENTIAL 4
SOME PANASAS CUSTOMERS COMPANY CONFIDENTIAL 5
PANASAS HARDWARE DESIGN Shelf switch(es) PSUs Battery Backup Module ActiveStor shelf Director Blade Storage Blade COMPANY CONFIDENTIAL 6
OBJECT STORAGE ARCHITECTURE COMPANY CONFIDENTIAL 7
OBJECT STORAGE ARCHITECTURE Client Nodes Up to 300 MB/sec of NFS performance per DirectorBlade File Request Object Map and Capability Switched Network Parallel data paths Up to 1.5 GB/sec per Shelf Unix and Windows Client NFS and CIFS Metadata Managers Object Storage Devices COMPANY CONFIDENTIAL 8
INSTALLATION AND SCALING MADE EASY 1. Panasas Customer ID: <panasas supplied> * 2. Administrator Password: * 3. Enable PanActive Link: <optional yes/no> 4. SMTP Server: 5. Administrator Email: 6. Secure Web Proxy: <optional if used> 7. System Name: * 8. Default Router: * 9. Blade IP Range: * 10. Max. Ethernet Frame Size: 11. DNS Domain Name: 12. Primary DNS Server IP Address: 13. NIS Domain Name: 14. NIS Server Name: 15. Enable NIS Hostname Resolution: 16. NTP server name: * 17. Timezone: * 18. Enable NFS: <yes/no> 19. Enable CIFS: <yes/no> 20. Enable Vertical Parity: <yes/no> Enter the number of the entry (1-19) you wish to change, Ctrl-c to quit, or "save" to save these settings: [save] Submitting settings now... Settings were accepted. It will take a few moments for configuration to take effect Simple Installation Assign IP address, netmask and default route to primary DirectorBlade via serial connection Then answer the installation questions NOTE: The questions are relating to network connectivity and not storage configuration since there are no LUNs etc to configure. COMPANY CONFIDENTIAL 9
INSTALLATION AND SCALING MADE EASY Simple Installation Remaining blades obtain their configuration automatically from the configured DirectorBlade DHCP on Private port Automatic Online Provisioning Configuration is read from primary blades Automatic Software version matching Immediately serving data Automatic Capacity Balancing Target New Writes Migrate Component Objects in the background COMPANY CONFIDENTIAL 10
SIMPLE VOLUME CREATION Virtual Volumes Simple creation Free to use all available StorageBlades* Optional volume capacity quotas Optional user per volume quotas' Mechanism to distribute metadata workload Each volume is assigned to a DirectorBlade for metadata services Single Name Space global mount or traditional mount points /home /data /test /bench COMPANY CONFIDENTIAL 11
ENHANCED RELIABILITY THROUGH SOFTWARE Fault Tolerant for Storage Cluster Management DirectorBlades clustered together Automatic Volume Metadata Failover Mirrored transactions between DirectorBlades Intelligent StorageBlades scrub disks in background Repairing grown media defects, parity and object attributes Monitor the S.M.A.R.T attributes of the disks S.M.A.R.T errors can indicate a future failure Blade Drain Objects are migrated away from StorageBlades which have predicted failures, preventing reconstruction COMPANY CONFIDENTIAL 12
ENHANCED RELIABILITY THROUGH HARDWARE Redundant active/active power supplies and fans Built in battery module for power fail protection Cached data written to disk Blades shutdown gracefully Drive heads always parked ECC protected memory Redundant integrated Ethernet switches per shelf Active/Active or Active/Passive COMPANY CONFIDENTIAL 13
Aggregate MB/sec ADVANCED RAID FEATURES Advanced RAID Per File RAID RAID Layout is an Attribute Stored within the Object Automatic transition from RAID 1 to 5 without restriping RAID 6 is coming De-clustered RAID Two level RAID MAP, Stripe width and depth Parallel Reconstruction DirectorBlade Clustering Object Reconstruction Distributed Spare Space Scalable Performance Small File Large File RAID 1 Mirroring RAID 5 Striping Reconstruction BW 120 100 1G Files 80 60 40 20 0 1 4 8 12 # of Shelves (1 DirectorBlade, 10 StorageBlades per shelf) Enables optimum system growth and reconstruction COMPANY CONFIDENTIAL 14
CHALLENGES: AVAILABILITY Challenges: Availability Ref: "Storage Challenges for Petascale Systems Dilip D. Kandlur http://www.dtc.umn.edu/disc/resources/kandlurisw5.pdf COMPANY CONFIDENTIAL 15
CHALLENGES: AVAILABILITY Challenges: Availability Ref: "Storage Challenges for Petascale Systems Dilip D. Kandlur http://www.dtc.umn.edu/disc/resources/kandlurisw5.pdf COMPANY CONFIDENTIAL 16
VERTICAL PARITY Solves media error problem regardless of drive density RAID within an individual drive Improves on internal ECC capabilities Independent of horizontal arraybased parity schemes Seamless recovery from media errors by applying RAID schemes across disk sectors Vertical Parity Vertical Parity Horizontal Parity COMPANY CONFIDENTIAL 17
NETWORK PARITY Extends parity capability across the data path to the client or server node Eliminates Silent Data Corruption Enables End-to-End data integrity validation Protects from errors introduced by disks, firmware, server hardware, server software, network components and transmission Client either receives valid data or an error notification Network Parity Vertical Parity Horizontal Parity COMPANY CONFIDENTIAL 18
ACTIVESTOR PRODUCT FAMILY ActiveStor 14 80TB ActiveStor 11 40TB Entry Level Scalability (TBA) ActiveStor 11 60TB Balanced Performance & Capacity ActiveStor 12 40/60TB High Bandwidth and Metadata Performance High Capacity Includes SSD Technology for Mixed Workloads Launching 17.09.2012 COMPANY CONFIDENTIAL 19
ACTIVESTOR PRODUCT FAMILY ActiveStor 11 ActiveStor 12 Product Focus Balanced Capacity & Performance Highest Performance ActiveStor 12 40/60TB Read Throughput (MB/sec) 1,150 1,500 Write Throughput (MB/sec) 950 1,600 File Creates/Sec. per Director Blade (Metadata Performance) 4,260 6,250 Capacity (TB) 40 / 60 40 / 60 Cache (GB) 40 + 8 80 + 12 Architecture 64-bit 64-bit ActiveStor 11 40/60TB Balanced Performance & Capacity Highest Bandwidth and Metadata Performance High Availability Network Failover Optional Standard Link Aggregation No Yes Director Blade CPU Storage Blade CPU 1.73GHz dual core 1.73GHz quad core 1.30GHz single core 1.73GHz single core Note: based on a single 1+10 shelf. COMPANY CONFIDENTIAL 20
SNEAK PREVIEW OF THE ACTIVESTOR 14 COMPANY CONFIDENTIAL 21
INTRODUCING ACTIVESTOR 14 Intelligent, Unified, and Cost-effective SSD/SATA Storage SSDs accelerate metadata and small file performance 2 or 4TB hard drives deliver high streaming throughput performance Optimized for real-world, mixed file size workloads Highest Performance and best Price/Performance More than double the per-disk SPECsfs2008_nfs.v3 NFS ops/s of EMC Isilon* Scales to 1.4M 4K reads/s for high small file performance Continued highest bandwidth performance, scaling to 150GB/s 33% increase in drive density (4TB) with no impact on RAID rebuild times Easy to Deploy, Use, and Manage Automatic SSD/SATA tier eases setup and manageability Fully compatible with existing ActiveStor 11 and 12 systems ActiveStor 14 PanFS 5.0 PanActive Manager *Source: Published SPECsfs2008_nfs.v3 benchmarks. See slide 26 Performance: NFS V3 IOPS for detailed justification. COMPANY CONFIDENTIAL 22
ACTIVESTOR BLADE ARCHITECTURE Director Blade Storage Blade ActiveStor Appliance CPU, cache, network Orchestrates system activity Metadata services CPU, cache, data storage Enables parallel reads/writes Advanced caching algorithms Full Rack Switch Module Up to 83TB per 4U chassis Scalable to over 8 petabytes Up to 1.6GB/s per chassis Easy to install, easy to manage Low Total Cost of Ownership 830TB & 15GB/s per 40U rack 10GbE networking InfiniBand Router 2 option for IB connectivity COMPANY CONFIDENTIAL 23
ACTIVESTOR 14 VALUE PROPOSITION First to Market with Intelligent Use of SSD to Balance Performance and Cost Store all metadata and <60KB files on SSD Speeds small file performance, directory listings, file system responsiveness Strong at Both Throughput and IOPS Ideal for mixed workloads Up to 14,000 random 4KB file read IOPS per shelf 1.6GB/s streaming bandwidth per shelf Higher Reliability 30-50% faster RAID reconstruction rate means no rebuild penalty for larger capacity drives Higher Density per node 35% higher density than previous 60TB models Improved Storage Utilization Dual parity RAID overhead lowered from 11% to 3% First 12KB of every file stored in metadata ActiveStor 14 Storage Blade 120-480GB SSD CPU 8-16GB Cache 2-4TB HDD x2 COMPANY CONFIDENTIAL 24
BIG DATA DESIGN/DISCOVER CHALLENGE Mixed workloads require bandwidth performance and IOPS performance from a single storage system Even data sets for large file, throughput workloads are actually mixed workloads consisting predominantly of small files Many critical file system tasks mean heavy metadata workloads File system directory listings, data replication / backup, file system consistency checking, object RAID rebuild Source: Panasas analysis of file system data sets from customers and prospects (Jan-Aug 2012) COMPANY CONFIDENTIAL 25
PERFORMANCE: NFS V3 IOPS 1 Shelf of ActiveStor 14T: 20,745 SPECsfs2008_nfs.v3 ops/second with an overall response time of 1.99 ms from only 27 data drives! More than twice as fast on per-disk basis as Isilon s fastest system based on 10K SAS drives + SSD Approaches NetApp s fastest system based on 15K SAS drives + Flash on per-disk basis 2 Shelves of ActiveStor 14T: 41,116 SPECsfs2008_nfs.v3 ops/second with an overall response time of 1.39 ms, showing near-linear scaling from the single-shelf result. 1000 900 800 700 600 500 400 300 200 100 0 SPECsfs2008_nfs.v3 ops/disk ActiveStor 14T (7.2K 3.5" SATA + SSD) EMC Isilon S200 (10K 2.5" SAS + SSD) NetApp FAS6240 (15K 3.5" SAS + Flash) Panasas actually has a true scale-out system, but we are optimized for the enterprise; they're optimized more for Linux/high-performance computing (HPC) workflows. --Sam Grocott, VP Marketing, EMC Isilon Sources: Panasas and http://www.spec.org/sfs2008/. Panasas benchmark disclosure forms will be published by Sept. 17 at: www.panasas.com/sites/default/files/docs/panasas_activestor_14_sfs_results_1089.pdf. Isilon result: S200-6.9TB-200GB-48GB-10GBE - 7 Nodes, June 2011, 58586 ops/s with an ORT of 3.14, total of 168 drives and 43.1TB NetApp result: Data ONTAP 8.1 Cluster-Mode (4-node FAS6240), Nov. 2011, 260388 ops/s with an ORT of 4.8, total of 288 drives and 95.7TB EMC quote: http://www.theregister.co.uk/2011/10/28/isilon_vs_netapp/ COMPANY CONFIDENTIAL 26
PANASAS ACTIVESTOR Scale-Out NAS Appliance for Big Data Workloads Leading Performance that s Fully Parallel Bladed design allows capacity and performance to scale linearly to 8PB at 150GB/s and beyond! No in-band filer heads or hardware RAID controllers to constrain performance Easy to Deploy, Use, and Manage Tightly integrated system Set up or grow capacity in under ten minutes Single, global namespace High Reliability and Availability Object RAID with vertical parity and parallel RAID reconstruction limits exposure upon drive failure High redundancy in hardware and software ActiveStor 14 10 shelves, 830TB COMPANY CONFIDENTIAL 27
Thank You Chris Weeden Systems Engineer COMPANY CONFIDENTIAL 28