mersenneforum.org > Data Error rate plot
 Register FAQ Search Today's Posts Mark Forums Read

2016-01-04, 18:12   #89
chalsall
If I May

"Chris Halsall"
Sep 2002

235018 Posts

Quote:
 Originally Posted by henryzz If anyone has further suggestions/ideas for graphs give be a shout.
Those are /very/ interesting graphs.

Could you possibly label the FFT lengths on the top of the X axis?

The spikes of the errors right at the FFT boundaries is to be expected, but the spike at about 3.4 * 10^7 is quite intriguing.

2016-01-04, 18:25   #90
Uncwilly
6809 > 6502

"""""""""""""""""""
Aug 2003
101×103 Posts

278916 Posts

Quote:
 Originally Posted by chalsall The spikes of the errors right at the FFT boundaries is to be expected, but the spike at about 3.4 * 10^7 is quite intriguing.
There were a bunch of bad early results at ~33,219,000 and above (the 10 million decimal digit level). There were many people who ran tests that took a year in that range.

2016-01-04, 18:49   #91
chalsall
If I May

"Chris Halsall"
Sep 2002

13·773 Posts

Quote:
 Originally Posted by Uncwilly There were many people who ran tests that took a year in that range.
Interesting... And it now takes many of us less than a day.

But I hear what you are saying... The longer the run, the greater the chance of an error....

2016-01-04, 19:40   #92
henryzz
Just call me Henry

"David"
Sep 2007
Cambridge (GMT/BST)

2×2,969 Posts

Quote:
 Originally Posted by chalsall Could you possibly label the FFT lengths on the top of the X axis?
There is no good way of doing that really. The lines are too close and the numbers overlap.

2016-01-04, 20:02   #93
chalsall
If I May

"Chris Halsall"
Sep 2002

13×773 Posts

Quote:
 Originally Posted by henryzz There is no good way of doing that really. The lines are too close and the numbers overlap.
Put the labels vertically (along the Y axis) or at a -45 degree (polar) angle.

Yes, the lower ones may overlap, but some of the others might not....

 2016-01-05, 16:49 #94 Madpoo Serpentine Vermin Jar     Jul 2014 31·107 Posts Another view of error rates Rather than look at error rates versus exponent size, I figured I'd plot them over time, by month. Here's a graph where the 1st axis is the known bad % for that month and the 2nd axis is the % of unknowns. Note that where the date of a result is unknown, I lumped them all together into "1995-01-01" so those are all a single data point. It's interesting to see that where the % of unknowns is nearly zero, the error rate hovers around 5%. When the % of unknowns creeps upwards, the known error rate starts to drop off. My theory is that for the leading edge of first time checks, the ones that get double-checked may have been from people setting out to confirm results of their work and thus not really a random spot check. Resulting in a lower than normal rate of *known* errors. Also, for my data set I excluded my own triple-check work of any exponent below 33M for the simple reason that the large # of them, especially all the stuff below 2M, would have skewed the error rate as a function of time in a pretty bad way. Just imagine another 25,000 or so "good" results in 2015 and what that would have done to the perceived errors. One thing I got out of this was that I may embark on a mini-quest to clear the handful (40 or something) of unverified exponents from 2002 to the Feb 2006 or whatever, get those sorted. Attached Thumbnails
2016-01-05, 18:33   #95
chalsall
If I May

"Chris Halsall"
Sep 2002

235018 Posts

Quote:
 Originally Posted by Madpoo Here's a graph where the 1st axis is the known bad % for that month and the 2nd axis is the % of unknowns.
Isn't the 1st (X) axis the date while the 2nd (Y) axis the % bad? But this graph is indeed interesting.

The reason I was very interested in henryzz's most recent graphs is they /might/ have to do with the 768K FFT issue we're currently working. But your graph explores this question a bit as well.

Is there any chance you could produce a similar graph based on CPU architecture? Perhaps three dimensional; error rate, date and architecture.

I would be happy to DC any LLs submitted by Skylakes to help with the sample set.

2016-01-05, 20:10   #96
Serpentine Vermin Jar

Jul 2014

31·107 Posts

Quote:
 Originally Posted by chalsall Isn't the 1st (X) axis the date while the 2nd (Y) axis the % bad?
Fine, fine... the 1st and 2nd datasets. Excel calls it a secondary axis, but whatever. LOL

Quote:
 But this graph is indeed interesting. The reason I was very interested in henryzz's most recent graphs is they /might/ have to do with the 768K FFT issue we're currently working. But your graph explores this question a bit as well. Is there any chance you could produce a similar graph based on CPU architecture? Perhaps three dimensional; error rate, date and architecture.
It does have me a little intrigued, plotting error rates versus different things. I'd looked previously at error rates versus the app version but besides a few "home brew" builds that were consistently bad I didn't notice anything obvious, but then I didn't graph it either.

Going by CPU architecture could be interesting though. It's not always obvious in the data what type of system ran it but I may be able to roughly lump them into GPU/CPU groupings. The ambiguous ones would be manual submissions where the app version in the text is generic.

2016-01-05, 22:53   #97
chalsall
If I May

"Chris Halsall"
Sep 2002

235018 Posts

Quote:
 Originally Posted by Madpoo Going by CPU architecture could be interesting though. It's not always obvious in the data what type of system ran it...
You might find it interesting to "sniff the traffic over the wire" from the clients.

I don't know how much gets into the Primenet DB, but here's a redacted sample from one of my machines:

Code:
http://v5.mersenne.org/v5server/?v=0.95&px=GIMPS&t=uc&g=xxxx&hg=xxxx&wg=&a=Linux64,Prime95,v27.9,build+1&c=Intel(R)+Xeon(R)+CPU+++++++++++X3323++@+2.50GHz&f=Prefetch,SSE,SSE2,SSE4&L1=32&L2=3072&np=1&hp=1&m=3825&s=2499&h=24&r=4000&u=wabbit&cn=a3_bu2&ss=58054&sh=xxxx

2016-01-06, 03:40   #98
Serpentine Vermin Jar

Jul 2014

CF516 Posts

Quote:
 Originally Posted by chalsall You might find it interesting to "sniff the traffic over the wire" from the clients. I don't know how much gets into the Primenet DB, but here's a redacted sample from one of my machines: Code: http://v5.mersenne.org/v5server/?v=0.95&px=GIMPS&t=uc&g=xxxx&hg=xxxx&wg=&a=Linux64,Prime95,v27.9,build+1&c=Intel(R)+Xeon(R)+CPU+++++++++++X3323++@+2.50GHz&f=Prefetch,SSE,SSE2,SSE4&L1=32&L2=3072&np=1&hp=1&m=3825&s=2499&h=24&r=4000&u=wabbit&cn=a3_bu2&ss=58054&sh=xxxx
That looked kind of like the messages that come in when a client is updating the server with its hardware info. Normally they don't include all that much detail. Info on the L1/L2 cache size, CPU speed, it's latest rolling average, a rundown of supported features (AVX/FMA/etc).

It is possible in many cases to match up a result with the CPU and application that did the work. It's the manual submissions that lose a little of that (the CPU info, although the app type/version is typically in there). And some older software didn't even include that.

2016-01-06, 20:50   #99
Serpentine Vermin Jar

Jul 2014

31·107 Posts

Quote:
 Originally Posted by henryzz I found fft lengths within http://www.mersenneforum.org/showpos...2&postcount=35 This is only the old x86 and sse2 boundaries but it is better than nothing. Anything after that is probably going to be less clear anyway as I think there are more fft lengths these days...
Found this... pointed me to where in the source that info can be found.

http://www.mersenneforum.org/showpos...postcount=1149

It's helpful to see the different max breakpoints for SSE2 and AVX, especially with an experiment I'm doing comparing multi-workers and FFT sizes interacting.

 Similar Threads Thread Thread Starter Forum Replies Last Post ixfd64 Hardware 4 2011-04-12 02:14 S485122 PrimeNet 15 2009-01-16 11:27 GP2 Data 3 2003-12-01 20:24 dsouza123 Data 6 2003-10-23 22:26 GP2 Data 5 2003-09-15 23:34

All times are UTC. The time now is 09:24.

Mon Dec 6 09:24:07 UTC 2021 up 136 days, 3:53, 0 users, load averages: 2.26, 1.45, 1.32