2006-12-06

• Parallel Prim's Algorithm
• Issac's Rose translators
• put travel plans up on wiki

2006-11-29

• Dave: progress report for SC'06
  • Glen Otero from Terra Soft Solutions, responsible for PS3 based cluster.

2006-11-01

• progress report: Gengbin
• svn demo
• ext resesrachers
  • John Gunnels
    • handles optimization for IBM
  • Jose Moreira
    • manages the bluegene systems software.
    • worked with Ninja of Padua's group
  • David Kincaid
    • works on *lapack, does numerical analysis.
  • Micheal Heath
• other issues
  • Cell workshop: http://www.cs.utk.edu/~dongarra/cell2006/

2006-10-25

• progress report
  • Eric
  • wrote a interface that allows fast multi-cast in BlueGene?'s 3D torus map.
• progress report
  • sayantan
  • parfum
    • model based load balancing to try and balance the refinement portion of the adaptivity phase
    • worked with the import code.
• PPL website: Filippo
  • pointed out several problems eith the website
  • the FAQ is in need of some serious revisions
  • the sublevels of howtos need to click to open, click to close
  • the subleves become corrupted when used with a small browser.
• Ext. researchers: Celso
  • Horst Simon
    • authot of NAS benchmarks
    • does parallel algroithms
  • Padua
    • works on compilers for HPC
    • Analysis and optimization transforms
    • wrote the Polaris framework.
• other issues
  • svn/cvs

2006-10-18

• progress report
  • Celso
  • Paulino's group and parfum
    • Issac is looking at their TOPS data structures
  • access to NCSA, Teragrid, Argone's BG/L, XT3@PSC
  • NCSA petascale pre-proposal accepted, need more work?
  • bigsim project is written, not submitted
  • Faucets grant was accepted
  • petascale project is being developed.
• ppl website
  • Chao/Terry
  • recent publications are now updated.
    • need to add a filter to take out tech reports?
  • Doghouse phenomina with the website
  • Terry is still cleaning up the photos.
• External Researchers
  • Phillip Mucci
    • developed PAPI
    • good performance tools
    • can insert performance stuff into program at runtime! Wow.
  • Allan Snavely
    • maps - memory bandwidth measurements
    • lots of valgrind type stuff, but looking at performance instead of accuracy
    • has paper that says linpack is bad for benchmarks, once you look at the memory usage.
    • works at UCSD
  • Ron Minnich
    • does BProc
    • also produces guard, for parallel debugging
    • relative debugging? Looks at performance
  • Steve Scott
    • develpoed X1
    • works on adaptive supercomputing
• Other issues
  • sysadmin - Filippo - need to upgrade to 10.1 from 9.1, do it if you have the time.
  • Kale -workshops progress
    • need to clean up the charm tutorial
    • sharcnet == ncsa for canada

2006-10-11

• progress report: Dave
  • running on Cell on NCSA
  • anmd ran once on cell hardware?
  • optimzation (simulators only work on SPEs)
  • PMUP @ PCT'06, "Charm++, Offload API, and the Cell Processor"
• visitor: Terry Jones (LLNL)
  • doing research into the effects of deamon and noise on performance, periodicity/frequency of noise in processor load
• PPL website: Terry W.
  • group photo

2006-10-04

• progress report: Chee Wai
  • https://charm.cs.uiuc.edu/private/technical/progressReports/cheelee/CheeWai061004.html
• PPL website: Terry
  • next week, she'll be taking photos
• External resaerchers Chao Mei
  • J.(Ram) Ramanujam
    • works on parallel languages, complie-time/runtime support for multi-scale applications on hihg end systems
  • michael scott
    • works on transactional memory, syncronization mechanisms
    • programmgn language pragmatics

2006-09-27

• Presentation: Charm++ on Cell by Dave.
  • pros
    • provides alot of power (256 Gflops/second)
    • allows for both fine and course grained parallelism
  • cons
    • hard to program
      • want progammer to focus on application code, not Cell specifics
    • hard to debug,
    • portablility.
  • arch
    • one main element, with multi threading
    • 8 processors only with access to 256K memory
    • DMA to move system memory to and from local stores for 8 smaller processors
  • offload api
    • handle setting up DMA (programmer has to do this explicitly?)
    • functions to:
      • init close intercade
      • sed work requests
      • check for work request completions
      • make progress
    • work requests end and
      • call a callback
      • need to be polled
  • progress
    • phase 1: design architecture
      • get charm runnning on PPE
    • phase 2, get charm to run the spe
      • create offload api - done-ish
      • add hooks to charm++ so it can use SPEs
    • phase 3
      • make the cell processing transparent to programmer
      • requires dependency analysis on which entry methods are safe to port.
• Progrwss Report: chao mei
  • works on NAMD
    • wants to improve the memory usage in NAMD so that they can simulate bigger molecules.
    • NAMD is 10 years old ands sounds scary
  • works on bigsim
    • need to learn more about bigsim, does it simulate parallel processors? What?
  • misc
    • contributed to namd proposal
    • make talks page to be automatically generated.
  • future
    • bigsim
      • out of core performance improvement
    • namd
      • charmization
      • reduce memory usage
      • bad scaling on infiniband
• ppl website: gengbin
• ext researchers -abhinav, Chao
  • Robert Germain - molecular dymanics simulation. Heads Blue Matter group, competes with NAMD
  • Vujay Saraswat - concurrent constraint parallel programming languages, works on X10
  • marc snir - one of the MPI people
  • Jarek Niepolche - did global arrays and remote memory copy apis
• other issues - Sanjay- languages
  • pros
    • idea that one level of abstraction shouldn't dominate unless it is really general. Don't force things down the programmers throat.
  • weaknesses
    • lack of communication - need to talk more to the people with the framework needs.

2006-09-20

• Presentations: Abinhav
  • wrote something that patches binaries, injects code that makes gprof output
  • generated discussion: is statistical sampling faster than timer calls woven into the code?
• Presentations: Chao Huang
  • charisma is cool, I need to talk to Chao
• PPL website: Eric
  • w3c link checker -- cool!
  • http://char.cs.uiuc.edu/people/bohm/terascale/
    • need to fill in with your usage data.
• other issues
  • upcoming conferences
    • write something to remind people when paper deadlines come up.
    • pritish will write the application, test his sql muster.
    • ITPPS?
  • machine switchover
    • nfs incompatibilities between suse 9.1 and 10.*
    • going to help Filippo with the firewall stuff.

2006-09-13

• Presentation: An Evaluation of Global Address Space Languages: Co-array fortran and Unified Parallel C, Issac
  • Languages:
    • yesterday: HPF, Split-C C-Region Library
    • today: titanium, upc, co-array fortran
    • tomorrow: X10, Chapel
  • patitioned global address space
    • single global view of data
    • Programer Productivity
  • Co array fortran
    • simple fortran extension
    • multiple images of same program execute
    • all images execute all statements they reasch
    • communication may occur at co-array accesses
    • explicit syncronization mechanisms.
    • SPMD
    • explicit data decomposition or co-arrays
    • fixed number of threads
    • not explicit message passing
    • one side communication, no matching of sends and recieves.
    • easier to compile than HPF
    • if as good as mpi, then why use mpi?
  • Compiling CAF
    • allocate arrays at runtime in the beginning.
    • map co-array access to real communication mechanisms.
  • problems with strides
    • do it in user library, run into problems with compiler optimization
    • by caf compiler
      • pcking is easy,
      • unpacking is hard since it requires code on the reciever side
    • library routines
      • limited support
  • features necesory for performants
    • group puts into one big message.
    • non blocking communcication
    • procedure splitting
  • upc
    • spmd
    • explicit affinity or data ditribution
    • explicit synchroniztion
    • sequential or relaxed consistency
    • one-sided communication
    • multi-phase barriers
  • compiling upc
    • messgae grouping
    • split phase (get data before it's used.)
• PPl Website: Sayantan
• external researchers
  • Ian Foster - grid computing guy at Chicago.
    • made the strand languge, parallel prolog
  • Ed Seidel
    • Cactus - an open source problem solving farmework for mesh type computational science.
    • http://www.cactuscode.org/
  • OOPAR - neat ideas for data dependencies object oriented framework.
  • bruce henderson
    • good theory work in parallel algorithms
• other issues
  • publish 1000 processor data, because we can.

2006-09-06

• Efficient Execution of Tightly-Coupled Parallel Applications in Grid Computing Environments
  • good performance in grid environment, high latency, processors with different speeds.
• Filippo's progress report on ChaNGa?
  • gravity comsology stuff
  • particle based simulation
  • algorithm for each timestep

1. particles divided among chare elements
2. build a tree amongst the particles
3. Use Barnes Hut, get nlog n time.
4. data needed residing on other processors fetched
5. Cache Manager buffers all imported data.
   • working on multi-step methods, compute timesteps differently in some places.
   • trees need to be re0build after every node movement. Reuse old tree information
   • blue-gene runs, things seems to scale?
   • bugs!
     • data sorter and partitioner
     • periodic boundaries
     • load balancer
   • things to do (new students)
     • new physics?
     • better load balancer, cache manager destroys graph structure.
   • CharmDebug?, added new flags (didn't understand)

2006-08-30

1. presentation

• • Issac and Aaron
  • Sierra project, similar to ParFUM?
    • handles adaptivity
    • things you want to do per element:
      • interpolate
      • point lies inside or outside
    • sounds like they used a bridge pattern
    • array interface in c++ that uses Fortran ordering,
      • sounds like it offers the same capabilities as Blitz++, perhaps it's more lightweight?
    • Sierra has a framework that allows you to mix and match various solvers in a unified way. Neat! I'd like to take a look at their design.
    • Sierra uses h-adaptivity. Could h-adaptivity be used with memfem?
    • Messages:
      • good to be ale to more your code into the framework
      • 3D adaptiviity might be really difficult. 1. progress report: Lukasz
  • all the work done since march
  • comlib testing
    • strategy applies to an array, manages communication for Charm arrays.
    • I didn't follow this presentation.
    • I need to find out more about histogram sorting in parallel. In fact, I always seem to get excited about parallel algorithms. 1. PPL website: Aaron
  • Aaron redid everything so that the main pafe use CSS.
  • Apparently there are presentations every week with what people have done to improve the website.
  • I need to make a presentation on the advantages of wiki's 1. Other issues
  • subgroups, meetings

2006-08-23

Gengbin presentation -

• How to get many different processes running at once on one machine
• isomalloc uses mmap underneath. How does this function work?
  • maps a file to memory