Started own thread so I could post pics without getting yelled at for breaking someone elses thread.


Below are charts/pics which seem to verify an overall drop in granted credit/hour from the periods October 12-15 2006 and January 16-20 2007. I'm only showing charts for two computers since the others are exceedingly similar and a waste of bandwidth. I am however showing their summaries.

The periods where chosen because they were found on all puters, and would be comparable to eachother as they contain the same "mix" of wus.

The chart is split into two parts the left side being the older data, the right the newer data. I chose the exact same number of data points from the older data as I had in newer data so they'd even out on the graphs and from an "average" calculation. The summaries found below each chart are positioned so the old data summary is to the left, and the new to the right of it.








I'm showing the X2 5200 data, just so you can see it. It's too new to have old data.

What do you all think?

Too many 5.5.0 users leaving or changing Boinc versions? Project stamped certain wus with the wrong coefficient? I don't know.

It's my opinion that it's not drastic, still inline from a percentage basis. I don't know if they can nudge it up a notch, but it might be a good thing.

tony

Here's a chart of ALL my collected data(since new credit system) and a summary for the AMD64 2800.



doesn't look to severe

Note: that one big spike was an SRAMAN wu, they admitted to making a mistake with

[edit]the data in the previous graph for the 2800 represents the right 1/3rd of the above chart, you can see some spike were found in the October data which aren't always found in the graph, especially the first ones when the new credit system started, and the in fact, the data now looks more like the data collected when the credit system started.

This is the same chart with the SRAMAN removed in an effort to make it easier to see, also changed some formats for the same reason.



If you look at the result ID's you can make out where it changes for 41xxxxx to 42xxxxx, the 42 is where the October data starts, the 57xxxxx is the January data.

Couple of points, meant for general assistance and not nit picking. :)

1. Your sample numbers are a tad on the small side to make any statistically robust analyses, particularly your sample of 26 WUs. Is there any way you can get to a sample of at least 100 in each time period?

2. What was your reason for picking the two periods? Ideally, for statistical purposes, the period should have been randomly chosen (alternatively, you could do a population comparison, would require no statistical estimates then because you would move into a fixed-effect census, however findings would not automatically generalise to other users/other time periods).

3. Rather than a comparison of averages, which I assume are means, boxplot analysis focussing on medians would probably be a better approach, as it does not assume a normal distribution - the boxplot groups would also have the advantage in showing how "normal" your WU credit distributions are. I don't know if you have access to any statistics package (Excel doesn't do boxplots, which is probably a blessing), but any stats package worth its salt will, including R (freeware, open source stats package).

Just a thought - are you assuming the same WU credit algorithm throughout the process? I don't know how the credits are calculated but any kind of "calibration drift" could potentially explain your findings. This effect could be found in the situation where the WU creation process goes "out of step" with the credit algorithm.

---

the graphs are coming up as 2x1 pixels for me! Using IE7. Anyone else getting this?

---

Danny, I'm using IE7, and they look like I expected.

It's tooooooooo early to reply to Mangenta yet. Need to make coffee. That reply will require some thinking, so I'm going to get some mental lubrication.

Couple of points, meant for general assistance and not nit picking. :)

1. Your sample numbers are a tad on the small side to make any statistically robust analyses, particularly your sample of 26 WUs. Is there any way you can get to a sample of at least 100 in each time period?

I can't go back in time and do more wu, and agree with the sample size comment. Not much I can do, but give all I have. Another point is that screen width is only so big, once you add in a couple hundred points this graph starts to look like a big black rectangle with a funny sawtooth top. Making it unuseable as well.

2. What was your reason for picking the two periods? Ideally, for statistical purposes, the period should have been randomly chosen (alternatively, you could do a population comparison, would require no statistical estimates then because you would move into a fixed-effect census, however findings would not automatically generalise to other users/other time periods).

Several reasons:
I wanted the displayed charts to be comparable to one another, so I needed to find time periods where the majority of my computers all did work.
There is a limited number of dates to choose from. I collect data for all projects, so there are periods where NO rosetta work was done. Some computers are newer, some older, one has been under my control, then out of it, then back again, etc etc, so it's seems better to keep the periods and quantities matching.

3. Rather than a comparison of averages, which I assume are means, boxplot analysis focussing on medians would probably be a better approach, as it does not assume a normal distribution - the boxplot groups would also have the advantage in showing how "normal" your WU credit distributions are. I don't know if you have access to any statistics package (Excel doesn't do boxplots, which is probably a blessing), but any stats package worth its salt will, including R (freeware, open source stats package).

Sounds like you know more about this than I. I don't have anything but Excel 2002. You might wanna start collecting data, or I can email you mine and maybe you can display it better. If you go the the Boinc Help desk, click on my name, you can email me.

Just a thought - are you assuming the same WU credit algorithm throughout the process? I don't know how the credits are calculated but any kind of "calibration drift" could potentially explain your findings. This effect could be found in the situation where the WU creation process goes "out of step" with the credit algorithm.

Here's what my tables look like:


Essentially, I copy and paste from the results page/host to this sheet. I check the result ID and copy and paste the boinc version and app version, then copy and paste the wu name and number of decoys. I use a function to calculate claimed credit/hour, granted credit/hour, seconds/decoy, and granted credit/decoy. So, the "Al Gore ithm"(and I bet you thought he just invented the internet) remains constant. (and yes, I spend a hour or two EVERY day collecting this junk)(ok, sometimes it's 3-4 hours every third day, depends on the project, and how frequently they purge/delete results)

Here's what 188 samples look like (gotta pity those with low res displays) from ALL my data collected by my AMD64 X2 4800. Result IDs' 37xxxx to 38xxxx are September results, 41xxxx to 42xxxx are October, the brief bit about 3/4 the way across are a couple from Nov and mid Dec (note: it appears I'd reset the bios and no OCing resulted in lower benchmarks, thus lower claimed/hour. The 52xxxx are from January.

You can see where my benchmarks (claimed/hour) varies in roughly 4 groups (the same as the collection periods). From this it appears Granted Credit has gone down for the Jan period vs the others.

(snip)
For anyone other than the two of us following this thread, I'm taking this offline until I have done the analysis on the WU credit data. :)

Astro can post the results once we have them. :) )

Others are following, please keep us updated as warranted...

(snip)
For anyone other than the two of us following this thread, I'm taking this offline until I have done the analysis on the WU credit data. :)

Astro can post the results once we have them. :) )

Rut Roh, Magenta is checking my work. Feel like I'm in skoooo again.