Day after day the RAC experiment slowly continues but I must admit it is going in a direction I didn't expect it to go in the direction it is going. The original idea was that under certain defined circumstances it should be possible to predict RAC evolution. If a host is running 24/7 the awarded credit should increase linearly. The reasoning in the BOINC credit system is that equal work should be rewarded equally. When I started the experiment I thought of using the past credit earned to predict the future credit. That could then be used to predict RAC. However...
Let's take a look at the current data:
I'll briefly explain the data shown. The light blue series shows RAC values that were calculated by the project server. There is no history of this data so these are just samples from when I checked the host page. The individual WUs are stored for some time. Based on the credit granted I calculated the RAC myself. This is the orange series. The grey series show the difference between the two as a percentage. As explained in a previous article this result can be considered exact.
At two points in time I made a prediction based on an average of the last 10 samples. These are the dark blue and yellow line. As you can see both predictions are worthless. The basic idea failed.
Nonetheless I find this very interesting. It learns us that RAC is more complex than originally anticipated. Currently I'm trying to figure out why RAC is not following the theoretical exponential curve. Since RAC itself is the result of a pure mathematical formula it can only be influenced by the input parameters. There are two: time and credit. I also plotted the credit per time, the green dots. The green line is a 4 sample moving average. What is very noticeable here is that credit per time seems to be increasing.
I have been doing some brief research as to what could be the case of this and found several reasons for this on the BOINC wiki. Reasons might be found in host normalization but I have to dig further. If you happen to have ideas please share them with us.
I would like to expand the experiment with some extra datasets. I'm looking for:
- people who want to start BOINCing one project for a long time 24/7 so I can follow the data to make a similar graph as shown above. I'm especially interested in the RAC buildup for (almost) zero.
- people who have been BOINCing a project for such a long time that the RAC is 'stable'.
If anyone is willing to help feel free to contact me here or on slack.
