Download presentation
Presentation is loading. Please wait.
Published byOswin Curtis Modified over 9 years ago
1
Alok Choudhary Dept. of Electrical & Computer Engineering And Kellogg School of Management Northwestern University I/O and Storage: Challenges Moving Forward
2
3/15/2004@ANC2 I/O for a Petaflop system! Want 1 Terabyte/sec? Please buy 20,000 disks, each with 50MB/sec. Come up with $50m just for I/O Want to spin all of them fully? Buy a 20 MW power plant Want to stop computing for a second for I/O? Lose 1,000,000,000,000,000 operations Pray that disks don’t fail while doing something – or double the disks Want to Lock and Synchronize so you can write small fragment Are you joking?
3
3/15/2004@ANC3 Emerging Storage Hierarchy Exabytes Terabytes Petabytes ≈ 1 µs ≈ seconds ≈ 2–10 µs ≈ 1–30 ms ≈ 250 ms Main Memory MEMs Low Power DRAM High Performance Disk Holographic Memory Massive Arrays of Idle Disks Tape Silos Want to manage I/O – good luck!
4
3/15/2004@ANC4
5
3/15/2004@ANC5 Some Complexity Dimensions
6
3/15/2004@ANC6 A Typical System? Ack: Gary Grider (LANL)
7
3/15/2004@ANC7 Proactive – not reactive Self Learning Self managing Autonomic Active “What not how?” For Performance, power, load balancing, scalability How do we get there? Next 5 years of R&D by our team I/O System HW/SW moving forward
8
3/15/2004@ANC8 E.g., extend OSD concept for large-scale parallelism
9
3/15/2004@ANC9 Usage Pattern Immediate Use On-line analysis, Visualization, Data reduction Local parallel access model Future Use Relationships, subsets, affinities Analysis types, use of results Is there a way to specify, storage, use this information for effective management and optimizations
10
3/15/2004@ANC10 What we need to do moving forward Immediate Use On-line analysis, Visualization, Data reduction Local parallel access model Future Use Relationships, subsets, affinities Analysis types, use of results Is there a way to specify, storage, use this information for effective management and optimizations
11
3/15/2004@ANC11 Metadata (who defines and manages?) Application Level Date, run-time parameters, execution environment, comments, result summary, etc. Program Level Data types, structures Association of multiple datasets and files File location, file structures (single/multiple datasets multiple/single file) Performance Level I/O functions (eg. Collective/non-collective I/O parameters) Access hints, access pattern, storage pattern, dataset associations Striping, pooled striping, storage association Prefetching, staging, migration, caching hints Historical performance
12
3/15/2004@ANC12 Passive/Active Storage Passive Storage Traditional model of storage Store/retrieve files/data sets from storage Active Storage Storage and software provides Functions (user defined, e.g., analysis, compression) Fail-in-place capability (active storage can reorganize, remap or replicate data upon failures) Transparency of management a big plus for users Promising for large-scale usage Performance capability available in storage processors
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.