Java Grande June 2 to 3 2001 Stanford ********************************

C++ KAI
=======
Compilers C++ --> C

Standards conformance
Customer support
Debugging
Compilation speed
Price
Development environment
Optimization


SPEC is a waste of time
small market optimizations (like complex numbers) ignored
VLIW like Itanium will not get good performance on "mass market compilers"

Templates: C++ based functional programming that runs at compile time
	unexpectedly powerful
	slow compilation
	only work within a "semicolon"

C++ now beats Fortran

Research directions
	domain specific
		F90 is dense arrays, complex numbers
	Domain optimizations
		compile time as well as run time

Pre-processor C++ to C++ -- No good
Plug-in Optimizer
Program inspects itself at compile time
Compile time meta-objects
Global much harder than local optimization
Need a compile time programming model ****************
	parallel programming 

Java a simpler backlash to C++
C# opposite lash!

CCJ Collective Communication Java package -- Good work
	Thread groups not process groups
	good comparison with mpiJava -- Issues
		Algorithms
		RMI
		Serialization
	Only cluster not SMP

Lawlor/Kale Object Arrays
	C++ F90 -- not Java
	"mature" system Charm++
	support dynamic distributed objects
		home procesor holds forwarding address

My Session
Pugh: Address Java Memory Model
Shoja: Change scheduling of Green Threads to support REal time multimedia
Hudson: Clever Garbage Collector via copying while job running

After Lunch
van Reeuwijk: Scientific extensions (templates, multi dimensional arrays)
Bacon: Lightweight classes
Cappello: Internet Computing as execution of graphs
	overlap communication and computation

Maloney
	Porting TAU (Performance Framework) to Java
	Using SciVis and mpiJava
Bull
	Excellent Fortran C Java comparison
Dublin Group
	Phenomenological analysis of bytecodes in
	a) different applications
	b) different compilers

Chatterjee	
CFD #1: Java/Fortran is about 3
Ctran is C written like Fortran (Similarily Jtran)
CFD #2: Irregular -- not much floating point
	Java Collections too slow

Jackal DSM Good
	Amsterdam Roland
	Uses very efficient RMI -- Manta

Yukiho Sohda Good
	DSM is natural for Java
	JDSM pure Java -- use existing good conpilers

Fabian Breg
	Survey of Previous Work including Michael's
	Bit naive?
	Bad matrix multiplication analysis

Cappello to Program Committee ***********************************************

Group is a topic in JXTA?  *****
HLA
Internet Task Force -- has some higher level server effort

Parabon ***************************************************
Rendering
	If small data
Sequence Analysis
	BLAST database growing faster than Moore's Law
	Send part of database --  as compute proportional to N**2
	communication like N. Choose Optimal N
Protein Folding
	Need ensemble
	Takes 2 weeks to fold a small protein 
Regression -- exhaustive search

Canary Tasks
	Check if provider is returning wrong answer
	or lying about time taken (dollars proportional to this)
30 day test drive is free

Parallel Computing
	Problem is unreliability not latency

Dennis: use JXTA as a local peer group to 
Reynders -- transactional message passing to replace MPI
	as ASCI machines too big to be reliable


See PPT for Forum Meeting

Numerical Working Group 
Joe Darcy much less optimistic than Reynders
	Generics have taken several years and no opposition
	Operator Overloading and value classes have opposition
		support by SunLabs insufficient!

Jose Moreira has preprocessor 
Some interest in [[]] notation
need 0 dimensional multiarrays

Expenses
	$2.50 parking
	Personal Car
	2 Avis
	1 air
	JG Registration
	Hotel
9.30 am thursday to Monday 3pm