Transcript Title
Accelerating Time to Results KC ZHANG
Panasas Technical and Business Development Manager [email protected]
Leader in Parallel Storage Systems
Agenda
Panasas introduction Customer successes Panasas solutions
Slide 2
Panasas Availability
Panasas, Inc.
Panasas
Founded by Garth Gibson in 1999. First Customer Ship in 2003 The fastest supercomputer in the world runs Panasas
Slide 3
Primary Investors: HQ – Silicon Valley Market Focus: o o o o o o Energy Academia Government Life Sciences Manufacturing Finance Technologies: parallel file system and parallel storage appliance World wide support with over 25 global resellers
Panasas, Inc.
Partnering to meet customer needs
Application ISVs Resellers Standards Development Slide 4 Panasas, Inc.
Recognized Product Innovation and Excellence
NAS Magic Quadrant “Visionary ” Best HPC Storage Product Top 5 Vendors to Watch in 2009 Top Collaboration Between Government and Industry
Roadrunner,
Top Supercomputing Achievement
Roadrunner, Los Alamos National Laboratory
Top Supercomputing Achievement
Roadrunner, Los Alamos National Laboratory
8 Panasas Customers Win HPCWire Awards in 2008!
6 Panasas Customers Win HPCWire Awards in 2007!
10 Disruptive New Storage Technologies Promise Big Changes Panasas, Inc.
Slide 5
Slide 6
Panasas Powers RoadRunner
Panasas, Inc.
RoadRunner at a Glance
Slide 7 Panasas, Inc.
Petascale Red Infrastructure Diagram with Roadrunner Accelerated FY08
Secure Core switches NFS and other network services, WAN Nx10GE NxGE Archive Nx10GE FTA’s I B 4 X Compute Unit Site wide Shared Global Parallel File System (Panasas) 10GE IONODES r e e F a t T Compute Unit 4 GE per 5-8 TB 10GE IONODES IO Unit M y r i n e t CU Roadruner Phase 3 1.026 PF Roadrunner Phase 1 70TF CU Scalable to 600 GB/sec before adding Lanes Slide 8 1GE IONODES IO Unit M y r i n CU e t Panasas, Inc.
CU Lightning/Bolt 35 TF
Leaders in HPC choose Panasas
SWIFT ENERGY COMPANY Panasas, Inc.
Slide 9
The Common Themes
Slide 10
A. Very complex problems and simulations B. Very large number of files being used concurrently C. Very large number of concurrent users/servers D. Consolidating Users and Clusters on one storage system E. Any or all of the above
Panasas solves the most difficult storage problems while delivering very high reliability in an easy to use appliance-like package.
Panasas, Inc.
Breaking Through the Bottleneck
Clusters = Parallel Compute
Linux Compute Cluster
Parallel Compute needs Parallel IO
Linux Compute Cluster Issues Complex Scaling Limited BW & I/O Islands of storage Inflexible Expensive
Slide 11 Single data path to storage
Monolithic Storage
(NFS servers)
Benefits Linear Scaling Extreme BW & I/O Single storage pool Ease of Mgmt Lower Cost
Parallel data paths Panasas, Inc.
Panasas Parallel Storage Clusters
What is Parallel Storage?
The architecture for scale-out file storage NFS Clustered NFS File Server File Server File Server Parallel NFS Slide 12 NAS: Network Attached Storage Clustered Storage: Multiple NAS file servers managed as one. Good aggregate performance.
Parallel Clustered Storage: File server not in data path. Performance bottleneck eliminated.
Panasas, Inc.
Panasas Storage Cluster: Built on Industry-Standard Components
Integrated 10GE Switch Battery Module (2 Power units) Shelf Front 1 DB, 10 SB Shelf Rear DirectorBlade Midplane routes GE, power
Slide 13
StorageBlade
Panasas, Inc.
Performance and Scaling
DirectFLOW client o Standard installable file system o o Supports all common Linux flavors Support up to 12K clients Panasas DirectFLOW® data path DirectorBlade cluster o Divides namespace into virtual volumes o Allows metadata to scale (no bottleneck) Demonstrated scalable performance o 30+ GB/sec of sustained throughput from a single filesystem
Slide 14 Panasas, Inc.
Scalable NAS - NFS/CIFS
Scalable NFS/CIFS server o o o o Load automatically distributed across scalable DirectorBlade modules Scale to satisfy growing number of clients Any DirectorBlade module can access any file Slide in a new DB, instantly get more NFS ops/sec into the same data Access same data from any protocol o o Integrates non-Linux devices into system 2+9 configuration typically best for NFS. Balances CPU ops/sec with disk ops/sec
Slide 15 Panasas, Inc.
Total Time in Hours to complete the job
400 350 300 250 200 150 100 50 0 Panasas Other Vendor A Other Vendor B Data Set
•
23 Million Traces
•
139GB input dataset
•
234GB output depth migrated image gathers
•
247MB per depth slice, 970 depth slices
Throughput of Reads & Writes (MB/sec)
60 30 20 50 40 10 0 Panasas Chart Legend Read Rate Other Vendor A Write Rate Other Vendor B Data Set
•
23 Million Traces
•
139GB input dataset
•
234GB output depth migrated image gathers
•
247MB per depth slice, 970 depth slices
1400 1200 1000 800 600 400 200 0
Aggregate Throughput for 24 Nodes
Data Set
•
23 Million Traces
•
139GB input dataset
•
234GB output depth migrated image gathers
•
247MB per depth slice, 970 depth slices Chart Legend Aggregate Read Throughput Aggregate Write Throughput Panasas Other Vendor A Other Vendor B
Job Time Activity
Panasas Chart Legend Processor Waiting on Data Computation Other Vendor A Other Vendor B Data Set
• • • •
23 Million Traces 139GB input dataset 234GB output depth migrated image gathers 247MB per depth slice, 970 depth slices
ActiveScale Operating System
DirectFLOW ® Protocol o Provides parallel data paths for maximum performance PanFS™ Parallel File System o o o Distributed and parallel file system Block management hidden behind object storage interface File management distributed across metadata managers Designed to be managed by non-storage professionals ActiveScan Predictive Media Management o o Continuous sweeps of all data and disk media in the StorageBlade If discrepancies are detected the system proactively corrects the media defects Predictive Disk Management o Anticipates disk problems with automated, predictive failure analysis; data is moved prior to failure, to avoid reconstruction Real-time monitoring of client load generation o Identify performance bottlenecks among storage users
Slide 20 Panasas, Inc.
Horizontal Parity: Panasas ObjectRAID
Parity calculated and written to disk(s) o Any failed disk can be reconstructed from the remaining disks Panasas ObjectRAID is faster o Uses multiple RAID controllers to run in parallel (“Parallel Reconstruction”) Panasas ObjectRAID is more efficient o Reconstructs only user data versus every sector on disk 800GB Blade reconstructed in 31 minutes at Los Alamos National Laboratory!
Horizontal Parity Slide 21 Panasas, Inc.
Unique: Vertical Parity
Solves media error problem regardless of drive density “RAID” within an individual drive Improves on internal ECC capabilities Independent of horizontal array based parity schemes Seamless recovery from media errors by applying RAID schemes across disk sectors
Vertical Parity Slide 22 Vertical Parity Horizontal Parity Panasas, Inc.
Unique: Network Parity
Extends parity capability across the data path to the client or server node Enables
end-to-end
data integrity validation o Protects from errors introduced by disks, firmware, server hardware, server software, network components and transmission o Client either receives valid data or an error notification
Network Parity Vertical Parity Horizontal Parity Slide 23 Panasas, Inc.
Manageability: Single Global Namespace
Panasas removes artificial, physical and logical boundaries o Eliminates need to maintain mount scripts or move data
Cluster 1 Cluster 2 Cluster 3 Cluster 1 Cluster 2 Cluster 3 Single Global Namespace Cluster 1 Results Archived Files Cluster 2 Results Cluster 3 Results
Traditional Storage Networks Slide 24 Panasas Storage Cluster Panasas, Inc.
Automatic provisioning for easy growth
Online Provisioning o Configure One DirectorBlade and all others obtain their configuration via DHCP on private port o New Storage is seamlessly integrated into the system DHCP on Private Port Reading Config Setting IP Addrs Matching Versions Growth without limitations o Terabytes to Petabytes o Single seamless namespace
Single Seamless Namespace!
Panasas, Inc.
Slide 25
Manageability: Automatic RAID configuration
Per File RAID o RAID Layout is an Attribute Stored within the Object System assigns RAID level based on file size o < 64 KB RAID 1 for efficient space allocation o > 64 KB RAID 5 for optimum system performance Automatic transition from RAID 1 to 5 o No re-striping Two level RAID MAP, Stripe width and depth o Automatically optimizes stripe size
Small File RAID 1 Mirroring Large File RAID 5 Striping
Enables optimum system growth and reconstruction
Panasas, Inc.
Slide 26
Manageability: Dynamic Load Balancing
1 StorageBlade Capacity 2 StorageBlade Performance 3 DirectorBlade Performance
Biases new data objects to new blades Dynamically moves data objects from filled blades as needed Data objects striped broadly for performance Dynamically moves objects from “hot” blades Cluster design assigns new clients to least utilized DirectorBlades
Slide 27 Panasas, Inc.
Proven Panasas Scalability
Storage Cluster Sizes Today (e.g.) o
Boeing
, 50 DirectorBlades, 500 StorageBlades in one system. (plus 25 DirectorBlades and 250 StorageBlades each in two other smaller systems.) o
LANL
RoadRunner.100 DirectorBlades, 1000 StorageBlades in one system today, planning to increase to 144 shelves next year.
o
Intel
has 5,000 active DF clients against 10-shelf systems, with even more clients mounting DirectorBlades via NFS. Release 3.2 will allow them to deploy up to 12,000 clients against a single system.
o
BP
uses 200 StorageBlade storage pools as their building block o Most customers run systems in the 100 to 200 blade size range
Slide 28 Panasas, Inc.
Fast Deployment
Panasas Appliance Model o Deploy solutions in hours and days vs. weeks and months o Ireland's most powerful computer (#117 in the world) was installed in three hours and powered up in just one day, thanks to a rapidly deployable computing platform from Silicon Graphics and Panasas.
http://biz.yahoo.com/prnews/090205/sf67219.html?.v=1
Slide 29 Panasas, Inc.
ActiveScale 3.2 Released Sept 2008
Performance
10 GE switch => 50% improvement in shelf performance Multi-core client performance tuning Infiniband connectivity RAID-10 volumes to optimize N-1 workloads
Reliability
Complete HA feature set with addition of NFS/CIFS Fail over Industry leading data integrity with Vertical Parity and Network Parity
Manageability
Snapshots NDMP support for easy backups
Slide 30 Panasas, Inc.
Summary
Slide 31
Parallel storage provides high performance for faster survey turnaround and more complex algorithms o 10s of GB/s in production seismic processing data centers o 50% performance increase per shelf with 10Gb Ethernet Scalability to support more complex data acquisition and larger clusters o Deployed on a single shelf on survey vessels o 12,000 core clusters in production today o 4PB+ systems in production today Proven across the E&P industry o All major ISVs: Landmark, Paradigm, Schlumberger o Operating on 6 continents for Service Cos., NOCs, Majors and Independents
Panasas is proven to cost effectively increase processing throughput!
Panasas, Inc.
For more information, call Panasas at: 1-888-PANASAS (US & Canada) 00 (800) PANASAS2 (UK & France) 00 (800) 787-702 (Italy) +001 (510) 608-7790 (All Other Countries)
Thank You
张克诚
13701026265
Panasas, Inc.
Slide 32