Message boards :
Number crunching :
Keeping Us Up To Date
Message board moderation
Previous · 1 · 2
Author | Message |
---|---|
Send message Joined: 15 Nov 07 Posts: 31 Credit: 56,404,447 RAC: 0 |
so if you're thinking 'travis couldn't possibly be sleeping right now' there's a good chance you're wrong, even if it's 5pm EST :P ROFL Too cool, Travis. Keep up the good work, no matter what time of day it happens. We'll try to keep the flames under control. ;o) |
Send message Joined: 28 Aug 07 Posts: 20 Credit: 36,099 RAC: 0 |
Not sure how many some of the users that complained, Has actually ever crunched on a newly formed Alpha project ? I think I'm close to working with 50 projects over the last years, Most of them are Alpha's and Beta's as I like to help them get off the ground even with my one pc. Compared to alot of those before this one, This one gets a grade of A from myself, Especially when he is basically the only one trying to keep his project going and all with whatever limited knowledge he may have had beforehand. Just my opinion of course. G^R |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
Not sure how many some of the users that complained, Has actually ever crunched on a newly formed Alpha project ? I think I'm close to working with 50 projects over the last years, Most of them are Alpha's and Beta's as I like to help them get off the ground even with my one pc. Compared to alot of those before this one, This one gets a grade of A from myself, Especially when he is basically the only one trying to keep his project going and all with whatever limited knowledge he may have had beforehand. We've actually already gotten some interesting results. We've crunched through 5-10 asynchronous genetic searches (AGS) for two different sizes of volumes. We're comparing this data to how the AGS converges on the Rensselaer grid and BlueGene (results we presented in the eScience paper linked on the main page), and hope to have another publication out this month comparing Boinc, the RPI Grid and the BlueGene for the astronomy project. Before the server issues, the project was assimilating about 2-3 workunits a second, which is comparable to the large partition of the BlueGene we have used (1024 processors). So far what we've done is pretty promising :) Once we get the new binaries, we'll be able to test all different kinds of searches like particle swarm, simulated annealing and some others to see which work best for the astronomy application and which scale the best and are most resistant to the heterogeneity of BOINC, just to mention a few things. Nate should also be updating the code in a week or so with the "convolution" which makes the calculation a lot more complicated - but more accurate. Instead of testing the astronomy model to fixed star points, it's tested against star points that are probabilistic distributions, increasing the computation by almost an order of magnitude. How that works out will be really interesting to see :) I've talked to Nate and hopefully he'll be posting a bit more about how the convolution works and what it'll be doing. |
Send message Joined: 12 Nov 07 Posts: 2425 Credit: 524,164 RAC: 0 |
Great Update!! |
Send message Joined: 8 Oct 07 Posts: 289 Credit: 3,690,838 RAC: 0 |
Great Job Travis from an ATA'ers(Alpha Testers Anonymous)perspective. Keep up the good communication you have started the last week or so :) What I would like to know is how the Boinc results matched the other 2 runs (RPI Grid and BlueGene) Were the results statistically acceptable? Which run might be termed canonical? Now that you have 3.Will you be reproducing Boinc results on the other 2 to see if they match on future runs? Or will Boinc be trusted to "stand" alone? I guess these questions should be posted on a science thread but you started this Travis ;) |
Send message Joined: 30 Aug 07 Posts: 2046 Credit: 26,480 RAC: 0 |
Great Job Travis from an ATA'ers(Alpha Testers Anonymous)perspective. Keep up the good communication you have started the last week or so :) Well all the searches have converged to the known optimal for the data set we're using, what we're really interested in is how fast the searches converge on the different computing environments. Since the searches are non-deterministic, what we've been doing is taking a group of them and doing an average of the population fitness's min, median, average and maximum for each generation. As the search progresses these all converge to the optimal value. So far, we've gotten the bluegene to converge the fastest. Right now we've got to crunch through the output generated for the boinc searches to see how it compares. |
©2024 Astroinformatics Group