The Sharing and Training of HPC Resources at the University of Arkansas Amy Apon, Ph.D. Oklahoma Supercomputing Symposium October 4, 2006
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Outline of Talk HPC at the University of Arkansas –Current status A New Mechanism for Sharing Resources –AREON HPC Training –New course delivery via HDTV collaboration with LSU Collaboration opportunities and challenges –GPNGrid and SURAGrid –Resource allocation issues
Amy Apon, Ph.D. ● University of Arkansas ● October 4, High Performance Computing Resources at the University of Arkansas Red Diamond supercomputer –NSF MRI grant, August, 2004 Substantial University match Substantial gift from Dell –First supercomputer in Arkansas Number 379 on the Top 500 list, June, node (256 processor), TFlops
Amy Apon, Ph.D. ● University of Arkansas ● October 4, More Resources Prospero cluster –30 dual processor PIII nodes –SURAGrid resource Ace cluster –4 dual processor Opteron –Our entry point to the GPNGrid/Open Science Grid Trillion cluster –48 dual processor Opteron –Owned by Mechanical Engineering –About 1TFlop
Amy Apon, Ph.D. ● University of Arkansas ● October 4, How are we doing
Amy Apon, Ph.D. ● University of Arkansas ● October 4, We are seeing research results Computational Chemistry and Materials Science (NSF) –New formulas for new drugs –Nanomaterials –Chemistry, Physics, Mechanical Engineering –Over 95% of our usage is in these areas
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Research results in other areas, also Multiscale Modeling DNA Computing Middleware and HPC Infrastructure –Tools for managing data for large-scale applications (NSF) –Performance modeling of grid systems (Acxiom)
Amy Apon, Ph.D. ● University of Arkansas ● October 4, We have done some significant upgrades For the first year we used SGE on half the computer and half of the computer was self- scheduled PVM jobs LSF scheduler installed May 2006 About 60 users, about 10 very active users
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Thanks to LSF, we are busy LSF Daily Pending Parallel Job Statistics by Queue (jobs waiting)
Amy Apon, Ph.D. ● University of Arkansas ● October 4, And jobs have to wait LSF Hourly Turnaround Time of Normal Queue
Amy Apon, Ph.D. ● University of Arkansas ● October 4, We have something exciting to share
12 Fayetteville Magnolia Russellville Fort Smith Arkadelphia Conway Monticello Pine Bluff Jonesboro Little Rock TULSA DALLAS MEMPHIS MONROE 25-July-2006 AREON Arkansas Research and Education Optical Network We ARE ON !
13 The first bond issue (last fall) failed Governor Huckabee of Arkansas granted $6.4M (PI Zimmerman) MBO loop between Tulsa and Fayetteville fiber is in place, network hardware is being shipped The campus (last mile) connections are in progress All is on target for a demo to the Governor on 12/5/06! Fayetteville Magnolia Russellville Fort Smith Arkadelphia Conway Monticello Pine Bluff Jonesboro Little Rock TULSA DALLAS MEMPHIS MONROE AREON Arkansas Research and Education Optical Network
Amy Apon, Ph.D. ● University of Arkansas ● October 4, AREON Arkansas Research and Education Optical Network This fall, Uark will have connectivity to Internet2 and the National Lambda Rail The bond issue is on the ballot again this coming fall If it passes then the other research institutions will be connected to AREON We hope this happens! The timeframe for this is about a year and a half
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Opportunities for collaboration with OneNet, LEARN, LONI, GPN, and others
Amy Apon, Ph.D. ● University of Arkansas ● October 4, A Demonstration Application High Performance Computing New course in Spring 2007 In collaboration with LSU and Dr. Thomas Sterling –We are exploring new methods of course delivery using streaming high-definition TV –We expect about 40 students at five locations this time –Taught live via Access Grid and HDTV over AREON and LONI, … –A test run for future delivery of HPC education
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Collaboration via GPN Grid –Active middleware collaboration for almost 3 years –GPNGrid is in the process of making application as a new Virtual Organization in Open Science Grid –Sponsored by University of Nebraska – Lincoln, includes participants from Arkansas, UNL, Missouri, KU, KSU, OU –Hardware grant from Sun and NSF provide 4 small Opteron clusters for the starting grid environment –Applications are in the process of being defined
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Collaboration via SURA Grid Uark has a 30-node Pentium cluster in SURAGrid Some differences with GPN –CA is different –Account management, discovery stacks are different –AUP policy is different - SURA Grid applications are increasing. Uark can run coastal modeling and is open to running other SURA applications
Amy Apon, Ph.D. ● University of Arkansas ● October 4, More Collaboration Mechanisms Arkansas is participating with the recently awarded CI-TEAM award to OU, PI Neeman –Will deploy Condor across Oklahoma and with participating collaborators LSF Multicluster provides another mechanism for collaboration AREON will give the University of Arkansas great bandwidth
Amy Apon, Ph.D. ● University of Arkansas ● October 4, UofA Current HPC Challenges We have some I/O infrastructure challenges –The system was designed to have a large amount of storage, but it is not fast Supercomputing operations –AC, power, and UPS need to be upgraded Funding models for on-going operations –How will basic systems administration and project director be funded?
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Collaboration and sharing bring a challenge Usage policies –How do you partition usage fairly among existing users? –How do you incorporate usage from new faculty? Current policy uses a fair-share scheduling policy. Dynamic Priority = (# shares) / (#slots*F1 + cpu_time*F2 + run_time*F3); Shares divided among largest users groups: chem, phys, others
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Collaboration and sharing bring a challenge Are max run times needed? –Almost everyone has them –Requires checkpointing of jobs which is hard to do with our current I/O infrastructure –Requires user education and a change of culture Are user allocations and accounting of usage needed? Your suggestions here
Amy Apon, Ph.D. ● University of Arkansas ● October 4, Questions? Contact information: