UncleAl Posted April 2, 2008 Report Posted April 2, 2008 Let's cut to the chase, http://www.mazepath.com/uncleal/uncleal2.zip That is everything you need for serial timing runs, plus a README file. 272K download. Does your iron outperform an AMD FX-55 at 36 min 4 sec? Quote
Morally.Corrupted Posted April 2, 2008 Report Posted April 2, 2008 I'll grab the file tonight - Will let you know my times .... Quote
Morally.Corrupted Posted April 3, 2008 Report Posted April 3, 2008 Ran the timer on Single unit P4-HL -though for some reason possible due to my own error my data output doen't resemble the sample data provided. l run it again on the slowest unit for you later 30000.000 10536230484820 0.99999999656345011230001.000 10537284108460 0.99999999488904498930002.000 10538337781848 0.999999992353169516 Quote
UncleAl Posted April 3, 2008 Author Report Posted April 3, 2008 The numbers are bang-on, as they should be for a given executable, radius/atoms/CHI. What was the difference in timestamps between time1.txt and time2.txt output files? An FX-55 does the second run, the three larger radius points, in 36 min 4 sec. The first file requires a (fractional) second and is only used to set a starting time for the second file. You don't need an external timer. Merely subtract the two timestamps. That gives better numbers - no other process adding to delays. From README in the ZIP, right at the top of the file, TIMING RUN: "timer.bat" script file runs executable static file "chibzh" through two radius intervals. No external math libraries are needed. The difference in timestamps is the execution time for the second run for three points: 30,000 A, 30,001 A, and 30,002 A radius. An AMD FX-55 requires 36 min 4 sec. :8^>) Quote
alexander Posted April 3, 2008 Report Posted April 3, 2008 on nix, you can always use the time command, uncle :) it records time just prior to execution, and calculates the difference after the execution completion.... try it, time cat /proc/cpuinfo eliminates the need for that section of the code all together..... you could also use a var to record unix time prior to execution and then just after the finish, the difference will give you run time in seconds... Quote
Morally.Corrupted Posted April 3, 2008 Report Posted April 3, 2008 I will compare the timestamps tonight and post them for you, I'll also hit a few other boxes with the timer as well too... Quote
UncleAl Posted April 3, 2008 Author Report Posted April 3, 2008 Thank you! No matter how you do the timing, it is only the timing that matters. Adding another 0.1 log(radius) extension to the graph would require ~63 days in the FX-55. It makes more sense to run in slack time in unused hardware or in a bored cluster. The fitted line may not be perfectly straight. The slight curvature, if any, only shows at large radii. http://www.mazepath.com/uncleal/bzhdense.png 78 atoms/unit cell, with hydrogens.http://www.mazepath.com/uncleal/bzdense.png 48 atoms/unit cell, no hydrogens. The hydrogens matter despite the slower crunch The payoff is a coin flip. If you contribute - after the experiment is run - either you get an Acknowledgement in the published paper (spacetime is chiral) or we don't tell anybody (no surprises observed). Every physicist knows what the answer must be! A similar question was asked Christmas 1956. On New Year's Day 1957 every physicist who knew what the answer must be was proven wrong, The Fall of Parity Question authority. Quote
Morally.Corrupted Posted April 4, 2008 Report Posted April 4, 2008 Ok ran the timer.bat file again on my AMD machine and checked the timestamps on the P4... P4 -07:14:3008:19:29 01:05:01 AMD21:06:2221:49:21 00:43:01 Quote
UncleAl Posted April 4, 2008 Author Report Posted April 4, 2008 P4 - 01:05:01AMD 00:43:01 You can appreciate the need for a bored cluster with the parallel processing version of the program. Is the AMD doing anything for a couple of months? Sigh. I can write up a serial script file for new radii with estimated execution times. A single P4 really isn't up to the task. Intel hafnium CPUs with 45 nm architecure will be interesting. Quote
Morally.Corrupted Posted April 4, 2008 Report Posted April 4, 2008 What I will do I setup the chibzh to run through MPI and see what sort of numbers the cluster itself will generate on the timer.bat file....will post results Quote
UncleAl Posted April 4, 2008 Author Report Posted April 4, 2008 You can try it but I doubt it is that simple if you want speed. My UK programmer worked almost four months transforming serial BigCHI into parallel CHIpir. Each value report is devastating to process throughput. Subtleties. He gave up, then had a dream with the solution. Source code length about doubles. Speed is everything. For a cluster run you need the big gun primed to fire, with documentation. Tell me how many processors can be grabbed and I'll mount a ZIP including a script file for execution. Quote
Morally.Corrupted Posted April 4, 2008 Report Posted April 4, 2008 You can try it but I doubt it is that simple if you want speed. My UK programmer worked almost four months transforming serial BigCHI into parallel CHIpir. Each value report is devastating to process throughput. Subtleties. He gave up, then had a dream with the solution. Source code length about doubles. Speed is everything. For a cluster run you need the big gun primed to fire, with documentation. Tell me how many processors can be grabbed and I'll mount a ZIP including a script file for execution. Well Currently, I have 4 - Alexander has graciously said he could commit 2-3 to the project provided I can run a suitable OS for PPC, etc.. but that's not an issue as BCCD has support in Beta 2.2.2 for PPC's - I can also try to enlist a few more so realistically within about a weeks time could have a 10 node cluster up and runnning. Quote
UncleAl Posted April 4, 2008 Author Report Posted April 4, 2008 Running serial BigCHI through MPI probably won't do it. Parallel-executing CHIpir source code is twice as large as that for BigCHI. Speed is important. We do everything to make the executable fast. Tell me how many CPUs are in the cluster and I'll mount a ZIP archive with a script file, the source code, and documentation. Parallel source code should be compiled for the cluster in which it will reside. Remember - compile long_double_precision - NO Microsoft compilers! Quote
alexander Posted April 4, 2008 Report Posted April 4, 2008 lol uncle, i can't even conceive a cluster of windows boxes.... it sounds so funny and wasteful, it would be like hiring an 85 metric ton John Deere excavator, to plant a year old oak tree in your back yard. Quote
alexander Posted April 4, 2008 Report Posted April 4, 2008 not to say that a load of windows machines can't make it into a cluster, but only in a distributed computing project... a dedicated cluster, that's a totally different beast, on it's own... Quote
alexander Posted April 4, 2008 Report Posted April 4, 2008 ok, i have 3 machines which i can donate for a little while to the project: i have 800mhz PPC boxa Netra T1my AMD64 3400+ box (desktop, but i would not mind, for a little while ;) )i may have a 1.0 ghz p3 sitting upstairs (my dad's PC, and i know he is not using that) Inte3rc3pt (aka morally corrupted) has800 Mhz boxanother slow boxP4HThis crazy AMD box I used to have a quad 933 p3 box (4 penguins was real nice)... but that's been given as a gift to someone :| let me see what i have kicking around the garage, i may have another machine or two that may need to be put together :) I may have another friend who may be able to donate his machine to the project, for a while anyways i also may have another friend who may be able to just give me his old box, but i have to call him ooh i wonder if squelch has anything.... I'll know for sure as it gets closer to monday... Quote
alexander Posted April 4, 2008 Report Posted April 4, 2008 Uncle, you guys ever think of buying time on the amazon cluster? Quote
Recommended Posts
Join the conversation
You can post now and register later. If you have an account, sign in now to post with your account.