Dynamic Adaptive Operating Systems -- I/O

Transcription

1 Dynamic Adaptive Operating Systems -- I/O Seetharami R. Seelam Patricia J. Teller University of Texas at El Paso El Paso, TX 16 November 2005 SC 05, Seattle, WA 1

2 Goals Present a summary of our ongoing work Solicit assistance from the greater community Hear comments (+/-) 16 November 2005 SC 05, Seattle, WA 2

3 Project Goal Enhanced Performance Generalized Customized resource management Fixed Dynamically Adaptable OS/runtime services 16 November 2005 SC 05, Seattle, WA 3

4 Project Challenges Determining What to adapt When to adapt How to adapt How to measure effects of adaptation 16 November 2005 SC 05, Seattle, WA 4

5 Project Methodology identify adaptation targets characterize workload resource usage patterns potentially profitable adaptation targets off line determine/redetermine feasible adaptation ranges off line/ run time define/adapt metrics/heuristics to trigger adaptation generate/adapt monitoring, triggering and adaptation code, and attach it to OS monitor application execution, assessing performance (gain) and triggering adaptation as necessary KernInst 16 November 2005 SC 05, Seattle, WA 5

6 Project Challenges Determining What to adapt When to adapt How to adapt How to measure effects of adaptation 16 November 2005 SC 05, Seattle, WA 6

7 I/O Schedulers I/O Subsystem Schedulers AS Deadline CFQ NOOP Schedulers/Controller... Storage System 16 November 2005 SC 05, Seattle, WA 7

8 Introduction Linux 2.6 provides four I/O schedulers: anticipatory (AS) deadline completely fair queuing (CFQ) noop Selection at boot time: one scheduler for all drives runtime: one scheduler per drive Default: AS or CFQ 16 November 2005 SC 05, Seattle, WA 8

9 Solution Automatic and dynamic selection of scheduler or scheduling policy (within a monolithic scheduler) 16 November 2005 SC 05, Seattle, WA 9

10 Example Determining What to adapt - I/O Scheduler When to adapt -?? How to adapt - Explicit/Implicit Selection How to measure effects of adaptation - Execution Time (??) 16 November 2005 SC 05, Seattle, WA 10

11 Example Adaptation Automatic and dynamic scheduler selection Proof of concept with two schedulers: one providing bounded latency on requests, the other providing fair allocation of bandwidth (presented at OSIHPA Workshop) 16 November 2005 SC 05, Seattle, WA 11

12 ADIO Scheduler I/O Subsystem Monitor Scheduler Selection DASS Request completion feedback CFQ P 1 P 2 P N Deadline R d R b W d W b Controller Storage System 16 November 2005 SC 05, Seattle, WA 12

13 Experimental Evaluation Enforcing Bandwidth Bandwidth (MB/s Deadline CFQ ADIO Draining time is a factor in reduced bandwidth Deadline CFQ ADIO Number of Threads (a) Sequential Sequential Reads Disk Utilization Disk utilization tiobench: 2 GBs data in 4 KB blocks with 2 to 64 concurrent threads Several scheduler swaps Bandwidth (MB/s Number of Threads (b) Random Random Reads Disk Utilization 16 November 2005 SC 05, Seattle, WA 13

14 Implicit Selection I/O Subsystem Adaptive Scheduler Deadline CFQ SCAN NOOP Adaptive Scheduler/Controller... Storage System 16 November 2005 SC 05, Seattle, WA 14

15 Goals -- Revisited Present a summary of our ongoing work Solicit assistance from the greater community I/O scheduling -- Interesting Applications VM -- Memory Hogs Other areas Hear comments (+/-) 16 November 2005 SC 05, Seattle, WA 15

16 Acknowledgements We thank: DOE (Grant # DE-FG02-04ER25622) and UTEP for financial support IBM Corporation for IBM Shared University Research Grant (SUR) SC 05 and Harvey Wasserman for creating and conducting this session You for your interest 16 November 2005 SC 05, Seattle, WA 16

17 References Seetharami R. Seelam, Jayaraman Suresh Babu, Patricia J. Teller, Automatic I/O Scheduler Selection for Latency and Bandwidth Optimization, Proceedings of the Operating System Interference in High Performance Applications (OSIHPA), St. Louis, MO, 17 September J. Axobe, Time Sliced CFQ I/O Scheduler, J. Axobe, Linux Block IO Present and Future, Proceedings of the Ottawa Linux Symposium 2004, Ottawa, Ontario, Canada, pp , July 21-24, tiobench: Thread I/O bench for Linux, Choosing an I/O Scheduler for Red Hat Enterprise Linux 4 and the 2.6 kernel, Borril, J., J. Carter, L. Oliker, D. Skinner, and R. Biswas, Integrated Performance Monitoring of a Cosmology Application on Leading HEC Platforms, Proceedings of the 2005 International Conference on Parallel Processing (ICPP-05), June Pratt, S., and D. Heger, IBM-Austin, Workload Dependent Performance Evaluation of Linux 2.6 I/O Schedulers, Linux Symposium, 2, July 2004, pp November 2005 SC 05, Seattle, WA 17

18 Thank You Questions / Thoughts? seelam@cs.utep.edu pteller@utep.edu 16 November 2005 SC 05, Seattle, WA 18