1 / 22

On the Role of Burst Buffers in Leadership-Class Storage Systems

On the Role of Burst Buffers in Leadership-Class Storage Systems. Ning Liu, Jason Cope, Philip Carns, Christopher Carothers, Robert Ross, Gary Grider, Adam Crume, Carlos Maltzahn Contact: liun2@cs.rpi.edu , chrisc@cs.rpi.edu , rross@mcs.anl.gov. Why do we need burst buffer?.

Download Presentation

On the Role of Burst Buffers in Leadership-Class Storage Systems

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. On the Role of Burst Buffers in Leadership-Class Storage Systems Ning Liu, Jason Cope, Philip Carns, Christopher Carothers, Robert Ross, Gary Grider, Adam Crume, Carlos Maltzahn Contact: liun2@cs.rpi.edu, chrisc@cs.rpi.edu, rross@mcs.anl.gov

  2. Why do we need burst buffer? • Modern HPC storage system is designed to absorb peak I/O burst resulting in • bandwidth waste. • This storage system design cannot keep pace with the growing data demands • as supercomputers evolve to the era of exascale. • e.g. statistics from ALCF Intrepid Blue Gene/P system shows that 98.8% of the time the I/O • system is at 33% bandwidth capacity or less, and 69.6% of the time the system is at 5% • bandwidth capacity or less. Aggregate I/O throughput over one-minute intervals on ALCF BG/P [1] [1] CODES:Enabling Co-Design of Multi-Layer Exascale Storage Architectures Project Proposal

  3. Storage System of Intrepid (BG/P)

  4. Software Stack on Intrepid

  5. Adding Burst Buffer to the Storage System Model • We build a parallel discrete event based leadership class storage system model[2]. • We propose a tier of solid-state disk (SSD) burst buffer integrating to existing I/O nodes. [2] Modeling a Leadership-scale Storage System, N Liu, C Carothers, J Cope, P Carns, R Ross, A Crume, C Maltzahn 9th International Conference on Parallel Processing and Applied Mathematics 2011 (PPAM 2011)

  6. Methodology Why Parallel Discrete-Event Simulation (DES)? • Large-scale systems are difficult to understand • Analytical models are often constrained Parallel DES simulation offers: • Dramatically shrink model’s execution-time • Prediction of future “what-if” systems performance • Potential for real-time decision support • Minutes instead of days • Analysis can be done right away • Example models: national air space (NAS), ISP backbone(s), distributed content caches, next generation supercomputer systems.

  7. ROSS: Parallel Discrete Event Simulator • Developed in ANSI C, API is simple and lean. • Using Jefferson’s Time Warp event scheduling mechanism. • Reverse computation. • Global virtual time algorithm exploits IBM Blue Gene’s fast barrier and collective networks. • Ross main page: http://odin.cs.rpi.edu/ross/index.php/Main_Page

  8. Study of Bursty Applications TABLE I: Top four write-intensive jobs on Intrepid, December 2011

  9. Burst Buffer Parallel DES Model

  10. Burst Buffer Model Parameters TABLE II: Summary of relevant SSD device parametersand technology available as of January 2012 Burst buffer latency model:

  11. Single I/O Workload Case • Using IOR benchmark • ( consecutive 4MiB write • requests) • Capture write time, excludes • open/close time • Full storage system uses 128 • file servers • Half storage system uses 64 • file servers Simulated performance of IOR for various storage system and burst buffer configurations

  12. Investigate Parameter Space Simulated IOR performance for various burst buffer configurations and I/O workloads Simulated application runtime performance for various I/O and burst buffer configurations

  13. Mixed Workload Experiments (1) Full storage system with burst buffer disabled. (Compute Node View) Jobs execution time = 5.5 hours Full storage system with burst buffer enabled. (Compute Node View) Jobs execution time = 4.4 hours

  14. Mixed Workload Experiments (2) Full storage system with burst buffer disabled. (Enterprise Storage View) Full storage system with burst buffer enabled. (Enterprise Storage View)

  15. Mixed Workload Experiments (3) Full storage system with burst buffer enabled. (Compute Node View) Jobs execution time = 4.4 hours Full storage system with burst buffer enabled. (Enterprise Storage View)

  16. Mixed Workload Experiments (4) Half storage system with burst buffer disabled. (Compute Node View) Jobs execution time = 5.8 hours Half storage system with burst buffer enabled. (Compute Node View) Jobs execution time = 4.4 hours

  17. Mixed Workload Experiments (5) Full storage system with burst buffer enabled. (Compute Node View) Jobs execution time = 4.4 hours Half storage system with burst buffer enabled. (Compute Node View) Jobs execution time = 4.4 hours

  18. Mixed Workload Experiments (6) Half storage system with burst buffer disabled. (Enterprise Storage View) Half storage system with burst buffer enabled. (Enterprise Storage View)

  19. Mixed Workload Experiments (7) Half storage system with burst buffer enabled. (Compute Node View) Jobs execution time = 4.4 hours Half storage system with burst buffer enabled. (Enterprise Storage View)

  20. Mixed Workload Experiments (8) Full storage system with burst buffer enabled. (Enterprise Storage View) Half storage system with burst buffer enabled. (Enterprise Storage View)

  21. Summary and Future Work • Burst buffer enables the following • Increase the application perceived bandwidth • Reduce the storage system while still achieving desirable bandwidth • Decrease application time to solution • Increase storage system bandwidth utilization • Absorb bursty I/O workload • Future work • Investigate the use of burst buffer on different tiers of the storage system • Enable the use of simulator framework to extreme scale based on current leadership class storage system model • Understand complex system components and facilitate the design through simulation

  22. Acknowledgements • Special thanks to Dr. Jason Cope; • Special thanks to my advisor Prof. Christopher Carothers, thesis advisor Dr. Robert Ross; • Thanks to Philip Carns, Kevin Harms, Gary Grider, Carlos Maltzahn, and Adam Crume. • Thanks everyone here today! • Questions?

More Related