if I have a computer farm, does it make sense to combine all results to one big results.dat and uses this as base on all computer? What are the advantages and disadvantages?
|You can do that, but if it gets too large it will take ages to load (and watch out for Muon1's auto-reduce of results.dat option).|
The advantage will be the possibility of faster evolution if the results.dat is synchronised between the computers rapidly. The disadvantage is that there will be less variety, so side-paths of evolution (like TOMAZ sometimes tries) won't be as likely to be found.
|If there are a lot of new computer, which runs only for a short time frame muon it would be better to use a synchronized results.dat. If there are more computer which run muon for a long time it would be better to not synchronize them. I will see how this will be for the boinc wrapper. Currently I synchronize the results. But the results.dat is 10MB now I need a tool on Unix to get the best 500 results of each lattice out of it as new results.dat.|
|yoyo, I think it is grat you will play with results.dat. There are some considerations:|
if there are lot of new computers which runs only for short time there is no need for such a large dat file (i think 500 results is large). If computers runs for short time there is little chans to get use of all gene pool from 500 results. By the time they get back again dat file will be new again so they will not use all of it again...Maybe 100 results file is OK (for one lattice).
Second thing is how to pick results for dat file. I am not sure taking best xxx results is the best thing. Problem is that we don't know what is the best. I suppose you have in mind best (highest) muon %. We are trying to optimize accelerator design to give highest muon% but question is if taking only highest results leads to that. Problem is that we can stuck at local extrema and we never learn about other possibilityes (if there is some other extremas). On my oppinion a better choice is to pick results with highest Mpts. High Mpts means that there is more particles in the system for a longer time. If such a run doesn't have good results it means something happenes with this particles due to bad design (they don't reach the end or they have wrong energy). So, if there are lot of particles and something is wrong with design it is reasonable to assume that that design have a good potential to be optimised. On the other hand if you take low Mpts design with high muon % there is nothing much to optimize and you stuck.
Unfortunately I don't have a computer farm to play with it so I put it here, maybe you or somebody else will play with it. On my oppinion good dat file(for one lattice) would be 80 results with highest Mpts, 10 with highest muon% and 10 random picked results (just to be sure not to get to narrow).
Stephen's dat files have another approach-they have all spectrum of results from 1 to 100 which is good for "general purpose" but question is if it isn't kind of too wide. That's why is good to try different dat file strategyes too (but you need reasonable ammount of computers for comparation with mainstream). Maybe you can provide all boinc users with your brand of dat file ?
Hope it will help and that it is understandable english. Damn spelling
|Thx Tomaz for the explanation. Does anybody has a tool which extract these results from my 12 MB results file?|
Should the extracted results be added to the samplefiles or send as results.dat to the clients?
Maybe we can make this experiment to extract the best Mpts and some randoms.