1. Is System X for Me? Cal Ribbens Computer Science Department
Lab. For Adv. Scientific Comput. & Applications (LASCA)

2. Outline System X configuration Programming model and environment
Problems most suited for System X
Next steps
Take-away messages

3. System X Configuration
A cluster: “multiple computers used as a single, unified resource.'' (Pfister, 1998)
1100 Apple G5 Xserves
CPU: PPC970, 2GHz, 64 bit.
Node: 2 CPUs, 4 GB RAM, 160 GB disk
Primary interconnect: InfiniBand
Secondary interconnect: Gigabit Ethernet
Peak theoretical speed: 17.6 TF
Linpack benchmark: 10.3 TF

4. Key Architectural Features
Distributed memory
Both physically and logically distributed
As opposed to SMP or DSM/VSM
Standard OS on each node --- unix
High-performance interconnection network
NFS over fast Ethernet for file i/o, data staging, etc.

5. Programming Model & Environment
Automagic compiler/IDE? Sorry.
New programming language? Maybe … if you are writing new code from scratch.
Commercial or community codes? Slowly emerging in some areas.
Existing languages (Fortran, C, C++) plus MPI? Most common. Possibly threads within a node.

6. MPI: Message Passing Interface
Programming model: multiple processes (usually each an instance of the same program), communicating by explicit message passing.
MPI is a de facto standard API.
Several implementations available for tcp/ip and for fast networks, e.g., Myrinet, InfiniBand.
Goals of MPI:
Portable, high-performance parallel codes
Support for writing parallel libraries

7. MPI Hello World include 'mpif.h'
integer myid, ierr, token, status(mpi_status_size)
call mpi_init (ierr)
call mpi_comm_rank (mpi_comm_world, myid, ierr )
if (myid .eq. 0) then
print *, "hello"
call mpi_send (token, 1, mpi_integer, 1, 0, mpi_comm_world, ierr)
else
call mpi_recv (token, 1, mpi_integer, 0, 0, mpi_comm_world, status, ierr)
print *, "world"
endif
call mpi_finalize (ierr)
stop
end

8. Code Development Environment
Standard editors, compilers, Unix tools. Work done on a “front-end” node.
Job startup: Batch system or command-line, e.g., “mpirun –np 4 myexecutable”
Debugging? Usual tools within node. Totalview for MPI.
Performance debugging? Tau

9. What will run well on X (and be worth the trouble)?
Large-scale computations: 100’s GB RAM, months of compute time.
“Regular” computations: where work and data is easily decomposed into many large, uniform chunks.
Dense vs. sparse matrices
Regular vs. irregular meshes
Molecular statics vs. molecular dynamics

10. What runs well? (continued)
Scalable computations
(work / communication) stays away from zero
Load stays balanced
Number of independent tasks scales up (and down)
Reasonable interprocess communication requirements:
Too much → communication will be bottleneck
Too little → not a problem; but maybe consider grid computing.

11. Next Steps Basic MPI short-course: Feb 17, 19, 24, 26.
Future LASCA short-courses likely
Consulting: grad student office hours or meet with LASCA faculty
Summer FDI on parallel programming
CS 4234 in F04, CS 5984 or 6404 in S05.
Refs:
Gropp, Lusk, Skjellum, Using MPI, (2nd ed.), MIT Press, 1999.
www-unix.mcs.anl.gov/mpi
research.cs.vt.edu/lasca

12. Last Comments HPC is hard: System X …
Models, machines, algorithms, software
Students should get credit
Collaboration is required
System X …
will make a huge difference for a few;
will make a big difference for many;
will be a success if we do great research, leverage the visibility, and build a community.