UVACSE Town Hall

1
UVACSE Town Hall

• University of Virginia Alliance for Computational
  Science and Engineering
• December 11, 2008

2
Agenda

• Who/what is UVACSE?
• Goals non-goals
• Metrics/strategy
• Outreach
• Tiger Teams
• Resource rationalization
• Issues
• Your thoughts

3
Who/what is UVASE?

• UVas response to the computational science
  revolution
• Formed after years of discussion by John Hawley,
  Mitch Rosen, Andrew Grimshaw,Rob Ribando, John
  OConnell, and many others.
• https//www7.seas.virginia.edu/itrtf
• Faculty led supported () by the administration
  and the VP-CIO
• Faculty allocation and policy board

4
People

• Director Andrew Grimshaw (CS)
• Staff
• Dr. Ed Hall
• Dr. Katherine Holcomb
• Mark Morgan
• Student tiger team members
• Karolina Sarnowska
• Duane Merrill

5
Goals

• Support sponsored computational science research
  in Virginia
• Change the culture of computation
• Education/outreach/curriculum
• Be the place in Virginia for HPC be a statewide
  resource
• Foster a multi-disciplinary ethos
• Assist ITC/UVA/the schools in planning, managing,
  and acquiring computational resources
• Rationalize high-end computing

6
Non-goals

• Build or manage a big machine for massive macho
  flops

7
Success Metrics

• How do we measure culture?
• Codes ported
• Cycles delivered
• Number of users of HPC
• Especially faculty
• Science done papers that result
• Bootcamps/classes attended
• Increase in research budgets associated with
  computational science

8
Strategy

• Outreach and education to change the culture
• Continue with boot-camps/short courses
• HPC, Unix/Linux, debugging, sci-viz, check
  pointing, data management, data mining, (Perl,
  Bash, Python, Matlab, Fortran) for the novice
• Work with schools/departments on computational
  science curriculum
• Tiger teams
• to assist with codes and culture
• move codes when appropriate to national resources
  to free up local resources
• Exploit technology to rationalize UVa resources
• PBS/Grids and integration into the Teragrid
• Inventory and integrate the large number of
  clusters on grounds need a carrot to get them
  to add to the stone soup
• Lab classroom machines for high throughput
  computing
• Work with ITC to solve the data problem
  2GB/user is pitiful. Direct access to home
  machine from any machine on grounds.
• Work with ITC resource managers to set
  computational science friendly policies.

9
Outreach Education

• Class in September/October
• Unix 22
• Fortran 26
• Adv. Fortran 15
• Matlab 15, 19, 20, plus a SEAS class
• Bash 13
• C 12
• January 2009 HPC bootcamp preparation underway.
• What sort of classes do your students, postdocs,
  and staff need?

10
Tiger Team Progress

• 10 responses
• Projects started
• Avik Ghosh parallelized, awaiting better
  algorithm
• Bill Walker 3X plus CUDA
• Nunemaker - done
• Ghyczy (Darden) vanished
• Haj-Hariri on-going
• Pasanak (English) - done
• Learmonth 9,000 jobs for his student
• Frank Diviney 10,000 jobs
• Tiger team proposals always welcome

11
Compute Data Resources
12
Queues/Clusters

• ITC Clusters Cedar (250), Dogwood (500), Elder
  (24)
• Contradictory goals
• High utilization (lots of flows)
• Low queue waits
• Long running jobs
• We have people who have put their equipment in
  the cluster and are suffering queue delays
• Solutions all have problems
• Shorten max job life
• Isolate owned resources
• Kill/suspend non-priority jobs when high priority
  arrive
• Dump jobs to CS (note we will only have a small
  number of nodes with very long reservations
  allowed)
• Get sequential jobs off of the cluster use
  Linux/Windows desktops
• Push people to national centers
• Accounting

13
Non centralized resources

• Department/research group clusters
• CS, Astronomy, etc.
• Desktop machines
• Linux Windows
• Public and private

14
Rationalization of resources

• Changed policy on clusters
• Utilization up but some problems
• Campus grid (XCG) is up and running
• 100s of desktop resources
• CS clusters and ITC clusters in the mix
• Support for high throughput computing, low-degree
  parallel jobs, and secure data sharing
• To utilize non-ITC clusters requires using XCG
• Three applications already, tens of thousands of
  jobs
• Expanding to VT next semester as part of XCG

15
XCG initial configuration
16
Accounting

• Historically CPU resources are free at UVA
• Different resources cost different amount some
  are more scarce.
• Queue controlled parallel machines are needed for
  some applications not others
• Some resource classes are much cheaper
• Desktop PCs (Windows, Linux, MacOS)
• There are more resource types than just CPU
• Storage, network, CPU, licensees
• Think of Amazons price model
• I believe we need a market model, where
  resources costs, different classes of resources
  cost different amounts, and people can sell
  their excess resources
• For example, Cluster resources .15SU/hr, Linux
  sequential .1SU/hr, Windows sequential .02SU/hr

17
Discussion

• How would you feel about an accounting and
  pricing model?
• What do you need to get your work done?
• Do you need storage more than cycles?