Streams
0) Connect Streams
1) Syncronous Read and Asynchronous Writes
2) isData available? is Emtpy?
3) How to encode EOS

Implmentations:
0) Independent unix processes for each page
   - write our own communications
   - same executable with different states
   - checkpoint and move required a respawn
1)  Single executable with multiple threads
    - good for 1 machine with smp
    - communications: streams as circular buffers
2) Multiple executables with multiple threads
   - 1 thread = page, 1 executable per machine
   - communication models?
3) MPI
   - communication is mainly taken care of, local optimized
   - don't use threads
   - assignment to machines
   - ganglia knows everything about everything

Execution model
0) rank = page , means several ranks put on the same machine
   - implement input numbers to distinguish input tokens
   - currently uses rexec to spawn processes
   - can we move the process between processors? checkpointing?
     serializing? what will MPI do for us?
1) one page per thread , multiple threads per process, one process per node
   - allows checkpointing and page movement
   - permits both spatial and temporal
   - implement a scheduling policy, sequentializing


-- 
Yatish Patel
yatish@cs.berkeley.edu