v3.0 WU Time Predictions
(Update)

Tweaking the CpF Values
When I posted the original article, I only had a few data points to base the work unit prediction on.  You can take a look at the graphs in the original article.  One correction I need to make on the graphs are the values of CpF (Cycles per FLOP) for the predicted CLI times for my PIII and Celeron.  Below is a chart of the values that were listed on the graphs and what they should have been. 

The actual number difference may not be that important, unless you are trying to directly compare my estimates to the estimates the reader may have come up with for their system. 

Ok now that is out of the way...The original estimates did a pretty good job of predicting the time for the client, but the predicted times were on the low end of the time spread.  The reason for this happening was because of the large angle range work units that I had processed.  There seems to be a trend for these work units to process significantly faster than the "predicted" times from Roelof's equations.  Why is that???  I have been assured that the actual TeraFLOP measurements for the large angle ranges were accurate, but it does appear that both Celerons and PIII's are more efficient in processing large angle range work units than mid angle range work units.  This lead to calculated CpF values for these work units to be a bit lower than those of mid angle range work units.  This skewed the average CpF measurement down a bit, and was not representative of the "average" work unit that you would get from the SETI@Home servers.   I have tweaked the values for the CpF and adjusted them up by 0.10 for both the Celeron and PIII CLI estimates, and that can be seen in the graphs below.  The adjustment of the CpF I did here was just an eyeball measurement, and probably could be tweaked up by maybe another 0.05....but it is close nonetheless :)

Difference in GUI and CLI processing???
If you noticed in the above graphs there was one thing I didn't touch upon which should be pretty obvious.  I have processed two work units on the CLI that are WAY off on the processing time estimations.  These two are on the PIII graph on the top, and are solid green circles.  What gives?  There has been quite some talk on the newsgroups about some work units taking WAY longer than the others, and that is what you are seeing here.  The "long" work units have low angle ranges (< 0.1), and are processing almost 2 hours longer than what is "predicted" by the equations that Roelof has come up with.  This is interesting.  I want to point you to the previous article and to take a look at the low angle range work units that I had processed with the v3.0 GUI.  Those work units processed with the GUI all fell within the range of the predicted times from Roelof's equations.  Is this something indicative of the work units themselves, or is there something different going on between the GUI and CLI clients?  To test this out, I reinstalled the v3.0 GUI, and reprocessed the same work unit (0.033 angle range), that I did with the GUI.  Here are the results for the same work unit done on both clients:

The work unit result in the table above is the red work unit on the graph.  The solid red dot is the result for the GUI, and the outline dot is the CLI time.  The first thing to notice and probably the most important is that the actual results for the work unit were the same if it was processed with the CLI or with the GUI client (I didn't include the gaussians, because the client doesn't do gaussian searches at this angle range.  A closer look at the result.sah file that each of these clients produced showed several small differences in the numbers generated, but these were usually just a variation up/down by 1 in the last digit (most likely not significant).  All of the spikes/pulses/triplets were found in the exact same places within the work unit.  The science is good. The processing and the processing efficiency seems to be different with these clients. 

How much different is the time for the CLI on these low angle range work units?  For the work unit detailed above, the time for the GUI client was spot on compared to the predicted time for that work unit.  For the CLI on the other hand, the work unit processed 2 hours longer than predicted.....in effect it took 54.7% longer than expected....and 34.7% longer than the same work unit on the GUI client.

Final Words
Now that there are some more data points for the graphs, we can come up with some conclusions about the run times of both the version 3.0 GUI and CLI clients.

1. Roelof's equations do a pretty good job for predicting the run times for the mid angle range work units on both the GUI and CLI version 3.0 clients.  There is some variability in the run times, but that is most likely due to the every day use of the computer in various web browsing, stats compiling and other stuff I do in between games of CounterStrike.
2. Large angle range work units consistently finish with run times below the predicted run times.  The CpF used for the equations are based on the "average" work units in which the clients will perform gaussian searches.  The gaussian searches cause a decrease in processing efficiency on average, which leads to an over estimation in run times for large angle range work units (which do not do gaussian searches).
3. For low angle range work units, the actual run times on the GUI are consistent with the predicted run times.  On the other hand, these same work units will have run times significantly longer than predicted for the CLI client.  I am not sure the reason why for this, but it appears that the clients do the same science, and will come up with the same returned results as the GUI client.  Something to remember though...the vast majority of work units that you will receive from the Berkeley servers will be mid angle range work units.  Only about 8% of the work units sent out have angle ranges < 0.1!

One last thing to take a look at with the CLI data...
Here I want to reintroduce a graph I had in the version 3.0 ramblings a couple of weeks ago:

What I want you to take a look at here is the version 2.70 beta results (green circles).  The v2.70 beta was known for having significantly long work units with angle ranges < 0.1.  It seems that the CLI client may be acting the same in terms of WU run times as the version 2.70 beta.  I just want to throw this out for discussion, since I can't really say that there is any connection other than the relative "shape" of the graphs and the WU time trends...  -zAmboni