Cluster Scheduling for Explicitly-Speculative Tasks

Cluster Scheduling for Explicitly-Speculative Tasks

Cluster scheduling for explicitly-speculative tasks DAVID PETROU December 2004 cmu-pdl-04-112 Dept. of Electrical and Computer Engineering Carnegie Mellon University Pittsburgh, PA 15213 Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy. Thesis committee Garth A. Gibson, chair Gregory R. Ganger Srinivasan Seshan Thomas E. Anderson, Univ. of Washington c 2004 David Petrou This research is sponsored by member companies of the Parallel Data Laboratory Consortium, by a National Science Foundation itr grant, and by the Army Research Office (contract daad19-02-1-0389). The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. The views and conclusions in this document are the author’s and should not be interpreted as representing the official policies or endorsements, either expressed or implied, of any supporting organization or the U.S. Government. ii · Cluster scheduling for explicitly-speculative tasks Categories and Subject Descriptors: D.4.1 [Operating Systems]: Pro- cess Management — Scheduling General Terms: Algorithms, Design, Performance Keywords: speculative scheduling, optimistic scheduling, cluster comput- ing, grid computing Imagine homemade sandwiches. iv · Cluster scheduling for explicitly-speculative tasks Abstract A process scheduler on a shared cluster, grid, or supercomputer that is in- formed which submitted tasks are possibly unneeded speculative tasks can use this knowledge to better support increasingly prevalent user work habits, lowering user-visible response time, lowering user costs, and increasing re- source provider revenue. Large-scale computing often consists of many speculative tasks (tasks that may be canceled) to test hypotheses, search for insights, and review potentially finished products. For example, speculative tasks are issued by bioinformaticists comparing dna sequences, computer graphics artists ren- dering scenes, and computer researchers studying caching. This behavior — exploratory searches and parameter studies, made more common by the cost- effectiveness of cluster computing — on existing schedulers without specula- tive task support results in a mismatch of goals and suboptimal scheduling. Users wish to reduce their time waiting for needed task output and the amount they will be charged for unneeded speculation, making it unclear to the user how many speculative tasks they should submit. This thesis introduces ‘batchactive’ scheduling (combining batch and interactive characteristics) to exploit the inherent speculation in common application scenarios. With a batchactive scheduler, users submit explicitly- labeled batches of speculative tasks exploring ambitious lines of inquiry, and users interactively request task outputs when these outputs are found to be needed. After receiving and considering an output for some time, a user decides whether to request more outputs, cancel tasks, or disclose new speculative tasks. Users are encouraged to disclose more computation because batchactive scheduling intelligently prioritizes among speculative and non-speculative tasks, providing good wait-time-based metrics, and be- cause batchactive scheduling employs an incentive pricing mechanism which charges for only requested task outputs (i.e., unneeded speculative tasks are not charged), providing better cost-based metrics for users. These aspects can lead to higher billed server utilization, encouraging batchactive adoption by resource providers organized as either cost- or profit-centers. vi · Cluster scheduling for explicitly-speculative tasks Not all tasks are equal — only tasks whose outputs users eventually desire matter — leading me to introduce the ‘visible response time’ metric which accrues for each task in a batch of potentially speculative tasks when the user needs its output, not when the entire batch was submitted, and the batchactive pricing mechanism of charging for only needed tasks, which encourages users to disclosure future work while remaining resilient to abuse. I argue that the existence of user think times, user away periods, and server idle time makes batchactive scheduling applicable to today’s systems. I study the behavior of speculative and non-speculative scheduling using a highly-parameterizable discrete-event simulator of user and task behavior based on important application scenarios. I contribute this simulator to the community for further scheduling research. For example, over a broad range of realistic simulated user behavior and task characteristics, I show that under a batchactive scheduler visible response time is improved by at least a factor of two for 20% of the sim- ulations. A batchactive scheduler which favors users who historically have desired a greater fraction of tasks that they speculatively disclosed pro- vides additional performance and is resilient to a denial-of-service. Another result is that visible response time can be improved while increasing the throughput of tasks whose outputs were desired. Under some situations, user costs decrease while server revenue increases. A related result is that more users can be supported and greater server revenue generated while achieving the same mean visible response time. A comparison against an impractical batchactive scheduler shows that the easily implementable two- tiered batchactive schedulers, out of all batchactive schedulers, provide most of the potential performance gains. Finally, I demonstrate deployment feasi- bility by describing how to integrate a batchactive scheduler with a popular clustering system. I have the fury of my own momentum. Bob, Fire Walk With Me Acknowledgements I thank Garth Gibson, my thesis advisor, for guiding my intellectual develop- ment with wisdom and patience. Garth taught me to ask the right questions and have defensible plans for answering them while giving me freedom to pursue problems interesting to me. Greg Ganger has been a second advi- sor, providing resources and dispensing advice. Both Garth and Greg have been supportive when crises caused me to take breaks. Tom Anderson was my undergraduate advisor at uc Berkeley and my research advisor in the Berkeley now Project. His words of encouragement, many years ago, con- stantly motivate me. I thank Srini Seshan for being on my thesis committee. I have been lucky to be advised by good people, in mind and heart. I thank the members of the Parallel Data Laboratory (pdl), especially Garth for creating and Greg for further promoting and developing this in- stitution, with its outstanding intellectual, personal, computational, admin- istrative, and economic resources. The following current and past members of the pdl Consortium provided support: 3Com, Compaq, emc, Hewlett- Packard, hgst, Hitachi, ibm, Intel, lsi Logic, Microsoft, Network Appli- ance, Novell, Oracle, Panasas, Quantum, Seagate, StorageTek, Sun, Veritas, & Wind River. Guests at pdl retreats expressed interest in and offered in- sights for my research. pdl staff members Joan Digney, Jennifer Landefeld, Karen Lindenfelser, & Patty Mackiewicz provided a positive work environ- ment. Other pdl and ece staff members supported my computing resources. Sharing 8208 Wean with Jason Flinn, Dushyanth Narayanan, & Sanjay Rao was often educational and always fun, despite music selection disagree- ments. Early on, Khalil Amiri was friend, elder gradsperson, and research collaborator. I profited from communicating with Sonya Allin, Mor Harchol- Balter, Miron Livny, Andy Myers, Jiri Schindler, & Steve Schlosser. Ex- changing ideas was a bonus to my friendships with Sourav Ghosh, John Grif- fin, Dushyanth Narayanan, David Rochberg, Craig Soules, Eno Thereska, David Tolliver, & Jay Wylie. My time as an eecs undergraduate at uc Berkeley was pleasantly passed with Will Chow, Daniel (‘danh’) Holliman, vii viii · Cluster scheduling for explicitly-speculative tasks John Milford, Sameer Parekh, & Ali Rahimi. Remzi Arpaci-Dusseau, Doug Ghormley, Brian Harvey, Carlo S´equin, & Amin Vahdat were inspirations. My Pittsburgh years have been happy, a time of varied experience and personal growth. I owe this being close to Dan Baselj, Julie Brick, Ben Feldman, Mark Lazarev, April Murphy, Dushyanth Narayanan, Hille Marika Paakkunainen, Jill Penman, Megan Schmidgal, David Tolliver, & Jay Wylie. Thanks for the existence of the 61C Cafe, where I was found holding court, courting, coding, writing, composing, and enjoying company. Baristas of note include Jason Bacasa, Keith Kaboly, Moshe Marvit, Nick Sarno, & Danielle Skoncey. Crazy Mocha’s Leah Loyd, Deanna Mance, & Dana Waelde generously hosted me during the final months. From California, my first best friends and bandmates I acknowledge: Ean Brown, Brian Gilmore, & Huy Huynh. Highschool friends shape each other, and I was glad to know Derald Brenneman, Joy(zelle) Davis, Sheila Salamipour, Kevin Stephenson, (the late) Stuart Tay, & Jason Thibodeau. My current Pittsburgh bandmates Hille Marika Paakkunainen and Mike Shanley provide an opportunity to play again. From the Music Department at Carnegie Mellon University, Nancy Galbraith, Natalie Ozeas, Marilyn Taft-Thomas, Donald Wilkins, & Colette Wilkins and the Dalcroze Eurhyth- mics faculty taught me and encouraged my musical aspirations. Closest to me are my late mom, my dad, sister, brother(-in-law), & niece, all from whom I receive overwhelming and unconditional love. Nothing in my life would work without them. My extended Italian and Greek families are also a source of

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    286 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us