Bad performance of Einstein@Home
23 Aug 2007 17:29:42 UTC
Topic 193071
(moderation:
Recently I joined to Einstein@Home (before I have been a memeber of seti@home project) and detected some strange performance problem. In correspondance with BOMIC manager I have 2 task and they should be delivered until 09.09.2007, but estimated to time to complete these task is about 71 hours. But as I remember when I've been in seti@home project their tasks completed much more faster (probably during 10 hours or more, but not 70 hours). What is wrong with Einstein@Home ? Probably I did not make some specific steps when I switched from one project to onother I do not know.
I have a very fast computer. In coorespondance with info from project I have
GenuineIntel
Intel(R) Pentium(R) 4 CPU 3.00GHz [x86 Family 15 Model 4 Stepping 1] [fpu tsc sse sse2 mmx]
and
Microsoft Windows XP
Professional Edition, Service Pack 2,
Log from BONIC manager
23.08.2007 20:46:10||Starting BOINC client version 5.8.16 for windows_intelx86
23.08.2007 20:46:10||log flags: task, file_xfer, sched_ops
23.08.2007 20:46:10||Libraries: libcurl/7.16.0 OpenSSL/0.9.8a zlib/1.2.3
23.08.2007 20:46:10||Data directory: C:\\Program Files\\BOINC
23.08.2007 20:46:10||Processor: 2 GenuineIntel Intel(R) Pentium(R) 4 CPU 3.00GHz [x86 Family 15 Model 4 Stepping 1] [fpu tsc sse sse2 mmx]
23.08.2007 20:46:10||Memory: 1023.23 MB physical, 2.40 GB virtual
23.08.2007 20:46:10||Disk: 39.06 GB total, 17.80 GB free
23.08.2007 20:46:10|Einstein@Home|URL: http://einstein.phys.uwm.edu/; Computer ID: 994229; location: (none); project prefs: default
23.08.2007 20:46:10||General prefs: from http://setiathome.berkeley.edu/ (last modified 2007-04-14 15:46:34)
23.08.2007 20:46:10||Host location: home
23.08.2007 20:46:10||General prefs: no separate prefs for home; using your defaults
23.08.2007 20:46:10||Reading preferences override file
23.08.2007 20:46:19|Einstein@Home|Restarting task h1_0496.40_S5R2__179_S5R2c_2 using einstein_S5R2 version 433
23.08.2007 20:46:19|Einstein@Home|Restarting task h1_0496.40_S5R2__178_S5R2c_1 using einstein_S5R2 version 433
Thanks
Bad performance of Einstein@Home
)
Also, as I know bouth projects use the same algorithm to process row data -Fourier transform.
The Einstein Wu-s are just
)
The Einstein Wu-s are just much bigger than the SETI ones.
RE: Also, as I know bouth
)
Einstein@Home, as far as I understand it, doesn't do ANY Fourier transform on the clients. The data that the clients receive is already fourier transformed,
(the abbreviation SFT that you will see in the output of Einstein@Home just after startup stands for Short(-Time) Fourier Transform)
the client then applies many filters, and computes a so-called F-Statistic (measuring how well the filters fits the data) and uses what's known as a Hough Transform to detect patterns in the data that indicate a candidate for a "signal".
A much better explanation can be found in the presentations concerning S3 results linked on the E@H homepage.
Back to the original question, every BOINC project uses differently sized workunits, for example ClimatePrediction.Net has workunits that take months (!!) to complete. Of course you get much more credit per workunit, then. It all depends on the underlying science.
CU
BRM
RE: ... What is wrong with
)
The others have given you the answer but I'd just like to add a couple of points. On average EAH results take about 5 - 10 times longer to crunch than Seti. This is normal and is certainly not a sign of a problem with the project or with your machine.
When you first start crunching for EAH, it is likely that the estimated time will be rather too large. If your initial time estimate says 70 hours, you may very well find that it only takes around 40 hours. BOINC will correct this estimate itself over time so it's not something to get too worried about. BOINC uses a parameter called "duration_correction_factor" (DCF) to make these corrections. The starting value is 1.0 and BOINC will gradually drop this value as it gains the experience of knowing exactly how long a series of results actually does take.
EDIT: You can keep track of what is happening to your DCF in two ways. You can see it on the website by viewing the information about your computer. The other alternative is to browse your state file (client_state.xml) in your BOINC folder. It is a text file and is easily browsed using Windows Notepad. Please be careful not to change anything though :). If you fully understand what you are doing you can change the DCF in the state file to shorten the time that BOINC would otherwise take to make the correction itself. Be careful as you could easily cause major problems if you made inappropriate changes. The responsibility is entirely yours :).
Cheers,
Gary.
RE: RE: Also, as I know
)
Agreed, and that's one reason why it's more feasible to do fixed server side credit scoring for EAH, since it's far easier to characterize the pattern matching work we do for the project.
With SAH, you have to chirp the data first, then FFT it, and then look for the signals, which are of an unknown pattern to boot.
Here all that 'messy' preliminary work is already done for us, and we then serach the preprocessed data for a single template frequency signature.
One other thing to keep in mind is the time length of the sample we search here for GW's is far longer than the now 86 seconds or so of scope time we look at for SAH.
Alinator
RE: RE: RE: Also, as I
)
Not really a single template frequency, but several frequencies within a small band of frequencies for each result, and the search is done for several thousands of different doppler-effect adjusted waveform templates. Seti at home knows where the radio telescope was pointed at at the time of recording, which implies certain doppler drifts from the movement of the earth around the sun. Einstein at home's observatories are omnidirectional, and we have to guess the direction from which a signal might come, adjusting for doppler drifts individually for each direction.
CU
BRM
Thanks for your help guys. I
)
Thanks for your help guys. I hope it behavior will be the same as you described.
RE: Not really a single
)
True, I was just generalizing (perhaps overly) the key points of the search strategy to emphasis the difference more.
However, even though SAH knows which way the telescope was pointing at the time of the recording helps with doppler drift in one regard, you still don't know what the shift might be from the motion of the actual signal source. That's why they chirp the data for a number of drifts as well (IIRC, a pretty big number at that).
In any event, somebody ends up burning a lot FLOPs on either project just getting the data ready to go before you can even start looking for something interesting! ;-)
In EAH's case, I'd hate to think how long a result might run for if the hosts did all the grunt work! Probably would need a lot more than 2-3 week deadlines.
Alinator