Meeting minutes
Purpose: release 2 presentation

Date: 11-1-07
Time: 2:00 - 3:00 P.M.
Location: Dr. Junk's Office JEB 324

Attendees: Dr. Junk, Frank Jones, Andrew Hattrup, Michael Krueger, Ryan 
Lease, Cody Tews, Dr. B.

Recorder: Frank Jones


	Michael: Explained release 1 features then introduced release 2 goals, pointed out the addition of the 
command line interface to the file-handler front end.


	Cody: "That's great"


	Michael: Showed a diagram of the updated operational model pointing out the addition of the parallel ecb 
execution path.  Next he introduced the updated architecture showing a diagram of our revised architecture next 
to our initial architecture.


	Cody:  "I misunderstood what you meant by fully utilizing the SPEs"


	Junk:	"Explain why you revised your approach..."


	Team outlined benefits of new design: Better usage of the bus, easier block management, better 
utilization of computational resources of the SPEs.


	Cody: Offered his opinion that the pipelined approach might lend itself better to a streaming model than 
the new approach.


	Frank: Pointed out inefficient bus usage and the dependancies that would cause the model to bottleneck 
in both ecb and cbc modes.
	

	Cody: Agreed the team had made the "best engineering decision"


	Michael: Explained how the program breaks a file into pages and pages are distributed to SPEs.  
Introduced the fact that we discovered we only had access to 6 of the 8 SPEs


	Michael: Reiterated our design approach pointing to the benefits we had already discussed and then 
outlined some of its limitations.  He pointed out that on smaller files the single threaded implementation was 
much faster and that we still needed to implement double buffering.


	Michael: Briefly summarized the new command line interface for the file-handler,... Cody was pleased


	Michael: Displayed some sample output from the script handler pointing out the added text indicating 
which algorithm was running.


	Ryan and Frank: Pointed out that this text was put in for debugging and left in for testing.  It will 
be 
modified or eliminated.


	Michael: Displayed test data


	Everyone: Lengthy discussion about results, latency and throughput for the parallel version were clearly 
better than those from the single threaded version for files larger than about 10 MB for latency, 4 for 
throughput.  


	Junk: Concerned by the inconsistency of the latency for the parallel version, suggested multiple runs and 
averaging the data to get a less "noisy" picture of performance.


	Frank:  Pointed to the second set of data (which was very consistent) and suggested the OS was the cause 
of the inconsistencies in the first set.


	Michael: "These results show that our algorithm performs better on lager files..."

	
	Junk:  "That's not what I'm seeing here....",  


	Team and Junk:  Review slides,...


	Junk: "Oh, OK I had my colors swapped..."


	Michael:  Summarized what we learned as a team, the unsigned long bit-width discrepancy and the 7th SPE 
being locked out were discussed.


	Junk:  Asked for more detail about the bit-width issue


	Frank:  Explained that on the Power Core an unsigned long is 64 bits, on an SPE it is only 32.  The 
original code was meant for an embedded environment where 16 bits is still common.


	Cody:   "yeah that's a nasty one"

	Michael:  Summarized what went well, stressing that the program works and that we didn't repeat release 1 
mistakes.  The simulator was an excellent tool and probably contributed to us discovering the SPE issue very 
quickly.


	Michael:  Discussed what didn't go well;  missed code freeze, people got sick,.. most significant was 
the VERY substantial amount of time that went into debugging this release due to the bit-width issue.


	Michael:  Explained that one of our mistakes was to create circular dependancies in the task list, 
we wont repeat that for release 3.  END OF PRESENTATION


	Everyone:  Discussed possible improvements in the design for release 3.  Dr. B. asked about hiding the 
startup latency by leaving the context of each SPE in-tact, perhaps using one of them as some kind of scheduler.  
The group discussed and came to the conclusion that the solution was to open multiple files from within the 
executable, meaning that the SPEs needed to be initialized only once.  Adding double buffering, larger I/O blocks, 
increasing a page size, and adding more instrumentation to the program were also discussed.


	Junk:  Told SEL people that we averaged 34 hours per team member last week, the current record for 
the semester.


	Dr. B.:  "They are ready to go to work!"


SEL EMPLOYEES EXIT....


	Junk:  Thanked the team for putting in the time necessary to make things work without needing to be 
prodded.  Stated that he wasn't as bothered by our missed code-freeze because of it. Emphasized that we need 
to make that freeze for release 3 because our time is more limited and our results must be more stable.


	Junk: Summarized the end of semester schedule and stressed that though it is usually a temptation 
for teams to slack off after the kind of effort we put in, we can't afford to do that, we don't have the time.


	Junk:  Asked if, other than the nasty bug we had enjoyed it


	Team:  Yes


	Meeting adjourned ...