Author: Charles Roberson
Date: 08:18:07 02/26/04
30 years ago, I spent time working for tobacco farmers. I dislike smoking of
any kind but it was the only job in town. There were several things that we did
that easily applies to parallel algorithms. These methods were all in how to
help out the slower people and getting the field primed faster.
All rows were not of uniform length -- the fields had curved boundaries.
So, here were some of the things we tried.
1) When the fastest person finishes, he helps the next person closest to
finishing. Then, those two help the next person closest to finishing and
so on ....
Variation: When the #2 person is finished the two leaders help split
their help across persons 3 and 4. Then the 4 split their
help across persons 5,6,7,8.
Once we had two or more helping the others "out". The
fastest person helps the person farthest behind and so on.
2) When the first person finishes, he helps "out" the person the most
behind. Then second person out helps the new person the most behind.
If the person originaly farthest behind is not the second "out" due to
the help. The two split efforts to help the most behind.
Now, which was the most effective and which did the farmers like the most?
#1 was the most effective and most liked by the farmer.
Why?
1) it made it very clear who were the slackers. (an issue with comps??)
2) the slackers were often taught better techniques and thus sometimes.
if they still didn't improve they weren't rehired.
3) it was better for morale -- the good received some amount of help but
so did the bad. Also, the best performers did not do too much extra
work.
Now, how does this apply to comps. Are there slackers? Yes, what about
distributed systems with different speed processors.
The first algorithm reminds me of Young Brothers Wait -- the last nodes
helped are the ones move ordering designates as least best.
The second algorithm reminds me of Bob's paper on DTS or any other work
stealing approach.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.