1. Scalability of Problem Solving Environments
2. and Wrap Up Session
   1. Ahmed Sameh Minnesota
   2. Wayne Joubert LANL
   3. Barry Smith ANL
   4. John Rice Purdue
3. Sameh Presentation
   1. *Scalable Algorithms for sparse problems adjust algorithms -- see METIS by Kumar for partitioning and matrix reordering
   2. *SPARSLIB
      1. -Iterative solvers not as robust as he would like
      2. -Can tailor to user's data structures
4. Barry Smith Presentation
   1. *Levels of abstraction
      1. -Application Interface
      2. -Math Interface
   2. *Data encapsulation(=hidden)
   3. *Use C as an example to remove compiler manipulation
      1. -objects carry with them routines to manipulate objects
      2. -in C specify functions by pointers to routines
   4. *State Encapsulation
      1. -Could use a global array -- common block but this is not good if recursive
      2. -So use a user supplied data structure
5. John Rice Presentation
   1. *Parallelism can be exploited at all levels from user to utilities
   2. *Data Parallelism in space is key
      1. -need to allow variable in space algorithms
   3. *Collatz conjecture (1970ish) -- time to solve PDE of same order as time to evaluate closed form solution
6. Joubert Presentation -- I
   1. *Amoco Cray LANL collaboration to develop parallel oil reservoir simulation system
      1. -GUI, visualization, geostatics,history matching. simulation itself
      2. -Want 100 times a Cray2
      3. -Must run overnight
   2. *Need scalability -- PC to MPP
7. Joubert Presentation -- II
   1. *Port to 128 node T3D -- initially 7-8 mflops per node
      1. -drastic optimization gives 18 mflops/node with 1 to 2 man-months
   2. *2200 wells, 5 years, 5 million grid points
      1. -run 50 such cases by end of year
   3. *Hindrances
      1. -lack of convergence of architectures and their short lifetime
      2. -no clear standards for sparse linear algebra
      3. -would like standard sparse BLAS so vendor would automatically optimize
8. Questions from Audience -- I
   1. *Physics of compositional models makes load balancing critical not solver
      1. -however Amoco problem is not of this type
   2. *Smith with usual glee with pointers(!) says we will use unstructured meshes for everything to allow natural load balancing
   3. *General discussion of load balancing ensues but little/new deep thoughts
9. Questions from Audience -- II
   1. *Memory used could be a serious issue on the web as can't predict how much memory availabe
      1. -somebody claims (remarkably) that cache based machines run irregular meshs as well as regular meshs
   2. *Skjellum notes that pointers in C++ are class relative so can move from machine to machine
   3. *Incorrect discussion of multithreaded models for load balancing irregular problems
10. Overall Conclusions -- I
    1. *Hardware will change
    2. *Need to support multiple paradigms
    3. *PSE's will deliver scalable libraries
    4. *Need Examples of PSE's inside and outside PDE domain
       1. -Purdue had a table of some 20 PDE PSE's
11. Overall Conclusions -- II
    1. *PSE's are good but need study of software architecture and tools to build them
    2. *PSE's will build on PSE's
    3. *Can re-use PSE modules and PSE infrastructure
    4. *Standards for Sparse libraries