Notes for Discussion with Profs


> We are intrigued by your proposal.  There has been a great deal of work 
> in related areas from OS and DB students here at Berkeley.  What is 
> unclear to us is how your perspective from SCORE will lead you to new 
> design ideas. 

New Design Ideas:


> We are also concerned that you may have more than 1 
> semester worth of work here.

Evaluate the amount of work:

> More specific questins: What are you going to use for 
> comparision/evaluation purposes.

Applications?

> What algorithms are you going to use to 
> reduce communication between nodes, and avoid bottlenecks?

Nature of the dataflow:

- feed forward pipelines can utilize asynchronous send/blocking
  receive effectuively hiding the latency? This depends on the
  underlying implementation of the communication layer in MPI.

- blocking (batching) of data token may need to be employed to
  effectively amortize the send/receive operations.

- topology of the dataflow graph can be examined to reduce the number
  of cycles that are split between two or more nodes, and thus are
  subject to throughput reduction.

- Should communication become a serious bottleneck: use token output
  statistics obtained by profiling. Attempt to partition the design to
  minimize the communication between frequently communicating nodes.

> Prof. Joseph 
> also had some questions about your simulation plans that we should 
> discuss.

---------------------

The paper on River addresses:

- data availability
- communication intensive applications