CUDALucas Result Monitoring Thread

[Brain]

Purpose: Detect possible mismatch constellations by data analysis.
If you wanna join copy and add your results to the version specific code blocks. Improve table as you wish.

Overview:

CL version # Good DCs # Bad DCs
v2.00               1 : 0

Detailed:

CUDALucas v2.00:

Exponent # Initial FFT # Final FFT # Good DC # Bad DC # Mean Error # Mean Temp # Overclock (%) #    CL compile #  OS     # Non default options
27002813 #     1572864 #   1572864 #       X #      O #       0.05 #      85°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -t

FFT size selection can be done automatically or manually. If automatically the initial and the final fft length can differ. You should monitor the first 10000 iterations. If manually and fft length too short CL will quit on roudoff error if -t param is set.

The following post is very useful for fft size selection:
http://www.mersenneforum.org/showpos...postcount=1188

[frmky]

For your records, here are 8 that I have recently done with CL 2.00, 4 DC's and 4 first time runs. The default FFT sizes were used.

Manual testing	45515927	C	2012-04-16 06:12	0.0	6f6aabebddedf2__	79.0207
Manual testing	45519853	C	2012-04-16 06:12	0.0	a69075acfc6fd5__	79.0275
Manual testing	45517631	C	2012-04-16 05:43	0.0	5eb2cee62ec335__	79.0237
Manual testing	45520627	C	2012-04-16 05:43	0.0	a5bb97c88caa0b__	79.0289
Manual testing	26225167	C	2012-04-08 21:32	0.0	9d12fa5014d145__	26.0795
Manual testing	26182463	C	2012-04-08 21:32	0.0	5e810118532d8f__	26.0370
Manual testing	26227759	C	2012-04-08 21:32	0.0	4ddaa561e08c0d__	26.0820
Manual testing	26216189	C	2012-04-08 21:32	0.0	4828875b3a8a29__	26.0705

[frmky]

And four more with the default FFT sizes:

M( 45518923 )C, 0x7e2c27cfabe928__, n = 2621440, CUDALucas v2.00
M( 45507911 )C, 0x05176de587eb74__, n = 2621440, CUDALucas v2.00
M( 45507437 )C, 0xdeef10d7c85b8e__, n = 2621440, CUDALucas v2.00
M( 45537893 )C, 0xf1d664c65a286d__, n = 2621440, CUDALucas v2.00

[LaurV]

One of the oldest mismatches of mine, still unsolved, the lowest in fact: 26070883. This one (quadruple check now) seems to be a hard nut to crack, and note that the last two tests were done by CL. I hope for the current (anonymous) tester that he is not using CL. Or if he uses it, then hopefully he will get the "good" residue, AND that residue not be in the list (i.e. last two residues including mine are bad). Otherwise he is wasting his time.

@kdgehman: what version of CL were you using? For me it was 1.61, and no triple check (the triple checking I introduced later, after "bad residues" started to popup, so now when I get a mismatch I will TC-it before reporting it, to be sure that my residue is the right one, see for example 26269081, for which I am sure that my residue is right, it was done two times with two different versions of CL, one of them being 2.0, and on two different cards).

[kdgehman]

@kdgehman: what version of CL were you using?

CudaLucas 2.00
fft length = 1474560

Do you have the interim residues from your test? It would be interesting to see how far they match.

[LaurV]

Originally Posted by kdgehman

CudaLucas 2.00
fft length = 1474560

Do you have the interim residues from your test? It would be interesting to see how far they match.

Not sure. I will check when I reach home. But it doesn't really make sense, the most of the errors are caused by hardware (overclock, overheat, bad video memory, i.e. pixel-errors, common for video cards). The real interesting would be if 2-3-more CL tests actually match, but p95 is different. Then we could say we spotted a software error in CL, and that would be bad. Like for example, assuming a p95 DC/TC for 26269081 shows that the initial residue was good, and my "double-double-checked" one (of which I am sure is good) is bad. THIS kind of situation would be much more interesting, uncovering software errors in CL. If you repeat a test 5 times and get 5 different residues, it could only mean that you have to lower the OC or use a better card. 99% of the errors are due to OC and O-heat.

[Brain]

CL version # Good DCs # Bad DCs
v2.01               0 : 0
v2.00               2 : 0

Detailed:

Exponent # CL vers # Initial FFT # Final FFT # Good DC # Bad DC # Mean Error # Mean Temp # Overclock (%) #    CL compile #  OS     # Non default options
27002813 #   v2.00 #     1572864 #   1572864 #       X #      O #       0.05 #      85°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -t 
27007399 #   v2.00 #     1572864 #   1572864 #       X #      O #       0.05 #      90°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -t

Feel free to add your results directly into this table. Unformatted posting are welcome, too. I will do my next tests with fixed fft length.

[Brain]

CL version # Good DCs # Bad DCs
v2.01               2 : 0
v2.00               2 : 0

Detailed:

Exponent # CL vers # Initial FFT # Final FFT # Good DC # Bad DC # Mean Error # Mean Temp # Overclock (%) #    CL compile #  OS     # Non default options
27002813 #   v2.00 #     1572864 #   1572864 #       X #      O #       0.05 #      85°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -t 
27007399 #   v2.00 #     1572864 #   1572864 #       X #      O #       0.05 #      90°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -t
29064461 #   v2.01 #     1769472 #   1769472 #       X #      O #       0.03 #      85°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -f 1769472 -t
29623213 #   v2.01 #     1769472 #   1769472 #       X #      O #       0.03 #      93°C #          0.00 # 4.0 / 2.0 x64 # Win7 64 # -f 1769472 -t

Feel free to add your results directly into this table. Unformatted postings are welcome, too.

I still think that mismatches may result from difference from initial and final fft length and the roundoff in between. But haven't examined yet. Will code restart from iter 0? So far everyting good. :-)