Discussion with Yatish, Norm and Yury


1. We want to port SCORE to millennium clusters.

2. What work needs to be done?

   a. Modify tdfc back-end to emit "simplified" c

      1. rebuild streams 
	 a. learn about shared memory
	 b. message passing and ????

      2. change the pagestep to be called continuously until done

   b. SCORE scheduler itself:
      1. remove the simulator 
      2. rewrite the interface between the scheduler and the simulator

	 a. what facilities to "pin" a process to a cluster node are
	    available?

	 b. are there ways to check point and move the process? 
	    GRANULARITY?

	 c. can we control our priority?

	 d. can we check the load any system? Do we engineer a policy
	    saying that if a node is more than X busy, then we think that it
	    does not exist.


3. The role of SCORE scheduler
   a. Maps the graph nodes onto computers in the Mill cluster.

      1. Are we doing 1-to-1 mapping?   streams to be MP only.

      2. clustering: many page-on-1 cpu? two levels of communication:
	 a. local comm: shared mem
	 b. global: mp or ???

	 How many pages should we schedule on 1 node (this depends on #CPU)?

      3. Do we move nodes around or are they assigned to one CPU for their
         lifetime (relocate)

   b. Instead of using CMBs to buffer intermediate results, use stream
      buffers (blocking reads)

   c. Temporal partitioning:

      1. Easier: Cut the graph into number of partitions == number of
         available mill nodes. Let OS sched take care of scheduling
         within the node. Use non-blocking writes to permit
         overlapping of compute and communication.

      2. Partition the graph into "schedulable" subgraphs (each
         subgraph has equal or fewer pages than available
         resource). Schedule each subgraph spatially and time-mux
         them.

      3. Understand the limitations in parallelism between (1) and (2)

3. Applications:

   a. we have and address the data set size
      1. jpeg codecs
      2. wavelet codecs
      3. mpeg encoder
      4. OTHER apps????
	 a. 3d rendering

   b. do we have uni-processor implementations?
   
   c. multi-processor?