P1) Is Linda good for scientific computing?

	- I take it as a premise that Linda *is* good for scientific 
	  computing, and do not feel it is my job to prove this.  
	  The contributions of my dissertation research do not 
	  rise and fall on this premise, anyway.

	- Others *do* feel that Linda is good for scientific computing.  
	  Simple model!

	- As far as efficiency goes, Linda is reasonably efficient:  

		* The costs of associativity have largely been 
		  removed by clever signature analysis and hashing.  

		* Besides, Linda applications rarely use the full power
		  of associativity, and especially not on more than
		  a few bytes per TS op.  

		* Besides, many Linda applications have a large granularity
		  of computation to communication.  BJORNSON SLIDE!!!

P2) Is checkpointing the only reasonable way to make scientific applictions ftol?

	- The jury is still out on whether providing ftol to Linda
	  is best via checkpoints or FT-Linda's state machine approach
	  or via transactions.  It is very difficult to estimate these
	  costs for Linda or message passing or other paradigms.

	- Checkpointing requires a global snapshot, which is non-trivial,
	  and failures can cause the entire program to roll back,
	  not just the process on the failed machine.

	- Taking checkpoints to disk is 1-2 orders of magnitude slower
	  than taking it to RAM....

	- Many FT-Linda applications can use scratch TSs for a lot
	  (think of dynamic bag-of-tasks), so this doesn't accrue
	  the per-AGS overhead of a replicated TS operation.

	- Combining 2-3 TS ops in an multicasted AGS costs little more 
	  a single distributed Linda operation or passing a message.

	- Non-replicated checkpoints are not a good idea, because you can have 
	  really long down times if just one computer fails (of course, the one
	  the checkpoint device is connected to).  And, of course, if you 
	  replicate checkpoints, they are going to be fairly expensive.