[wlonh]

reader50
Senior Member

Posts: 205
Registered: Jun 2000 posted 07-16-2000 07:09 PM ÊÊÊ ÊÊÊÊ ÊÊ
------------------------------------------------------------------------
I recently downloaded Motorola's LibMotoSh floating point library and decided to put it to the acid test using SETI@home. I copied my next work unit (entire SETI folder) and set the copy aside. After crunching the WU, I again copied the entire SETI folder as the Normal result. The original unsolved folder copy was substituted, the LibMotoSh extension was installed, reboot, and SETI was allowed to re-crunch the same WU using the LibMotoSh library. The final result folder was again copied as the LibMotoSh result. Then I compared the Normal result folder & the LibMotoSh result folder.

Configuration details: G4 350 Sawtooth, 64Meg, VM on, SETI used as screen saver only, no other applications running when SETI runs, screen set to go blank after one minute. I have tried turning VM off but got no speed difference. The main result was that there was no memory left over for anything else, so VM stays on. Will get lots more memory when the price comes down, I have a strong reluctance to pay double for the same product.

My normal average time runs about 7 1/2 hours at present, with some variation. The test WU ran longer than average. Screen shots compared:




Next, I compared the result files. The good results turn out to be in the Outfile.sah, where the client sends back the best result with plain english labels. The Result.sah file contains headers with the Outfile contents tacked onto the end. My Outfiles were identical with both methods except for time logging, and calculated spike power. I have copied out the spike power values below:

Solved Normal: power = 8.855247e+01
Solved LibMotoSh: power=8.855243e+01
LibMotoSh error: 0.99999954829
Time Normal = 8.15111111111 Hours
Time LibMotoSh = 8.7205 Hours
Time Difference: 1.0698541439 or about 7% longer.

General Conclusions

* System stability was not an issue. No odd behavior was noted.
* LibMotoSh looks like it cuts a few corners in the rounding department. The difference is extremely small, and would be significant only in repeated engineering calculations where exact accuracy is critical. I do not believe that this compromises the SETI@home results, so I see no ethical questions here.
* I suspect Apple has patched the standard SANE routines to take advantage of AltiVec when running on a G4. LibMotoSh, on the other hand, was last updated 5/31/1996 and could not have any AltiVec awareness. It would be interesting if someone with a G3 or earlier repeated my test to see if LibMotoSh takes longer on non-AltiVec systems.

On my G4, I have removed LibMotoSh. I chose to send in the Normal result.

[wlonh]

that is a repost , btw... reader50 did me no favors by not posting it here, the fora are not setup to transfer directly from the other fora to the seti forum so i had to yoyo a couple of times copying and pasting

[reader50]

Sorry, wlonh. I had run into LibMotoSh in the 3rd party forum, and my test was directed at LibMoto rather than SETI@home. The 3rd party forum seemed like the place to put the post.

btw: you killed my UBB formatting! For those who find the post here, the original version looks a little better. vanity and all that...

[This message has been edited by reader50 (edited 07-16-2000).]

[Bockie]

Well that certainly is a thorough examination of the results. Reader, thanks for the insight! I was debating on whether to use Lib or not; I now have my answer!

[wlonh]

of course, those with G3s may find different results... i do believe that i do

but i am not about to take the time to check it other than casual observation

i do not need the downtime it would take to run a test even though it is not that hard to do

I have a G4 500 with a gig of ram and before the libmoto extension, it was doing 1 unit in 4 hr 45 min plus or minus 5 min. After the libmoto extension was loaded, I never got below 5 hr 50 min on that box. I have removed the extension and now I am getting a unit in less tha five hours. However, my brother (Firewire) has installed it on his B/W G3 and is seeing improvements.

Moof