The GCP database has a small number of known errors due to hardware problems in a computer or an REG/RNG source. The database cannot be edited to exclude these errors, so they must be accounted for in analysis. Fortunately, they affect very little of the time we have been collecting data, but the list is presented here so that any analysis which covers an affected date can exclude the problematic egg’s data. A computer readable (CSV) file of egg errors on available on request. It is the best source to use for excluding bad data, and must be used for formal analyses.
Following are the larger chunks of bad data that should be excluded from any analysis.
- Egg 1005, Defective data from December 10 1998 to December 17 1998.
- Egg 1005, Defective data from October 31 to November 3 2002, with good data beginning at 13:25 PST = 21:25 UTC again (analyst may drop this egg on Nov 3). New OS and machine, trouble bringing system up, resulting in a mean of 255 for several hours, 17:25 to 22:19:50, on 20021031.
- Egg 106, Defective data on 18 November 1999.
- Egg 1025, Defective data from 9 November 1999 to 21 November 1999.
- Egg 1026, Defective data from 17 November 1999 to December 30, 1999.
Egg 1000, Defective data from 10:40:01 UTC 16-OCT-2000 until 07:39:58, 23-OCT-2000.
Running tests with substituted defective RNG.
Egg 28, Defective data from March 20 2001 to May 30 2001. Later discovered the bad data started earlier. Exclude from Feb 20 to May 30 2001.
Power supply replaced, data checked on 31st and are now nominal.
- Egg 28, Defective data from August 15 2002 to Aug 17 2002; variance ~28, indicating large bias corrected by the XOR. Power supply again, most likely. I will substitute a Microreg.
Egg 1101, Defective data on December 25, 2001.
Unknown cause, Mean is ~15 or 16.
- Egg 223, All data recorded from egg 223 must be discarded. It had an out-of-spec power supply and sent data with extremely small variance, largely determined by the XOR rather than the random source. The egg has been changed to an Orion, with egg ID # 1223.
Egg 118, from March 28 to May 19 inclusive, 2001, all data should be excluded. There may be none because 118 was down for change to new system, then reporting impossible trial values of 255 during this time because of a serial port problem. Good data begins on May 20 at 18:15.
The above supersedes an earlier note which said:
Defective data from May 20 to end of June, 2001.
Intermittent, unknown cause, possibly power deficit? We have established that likely cause was incorrect power from Mac serial port.
Egg 2222, Defective data from June 13 to June 15, and June 19, 2001
Brief shutoff at 14:15 to 15:00 on 13th, then Z’s of 20 and 30, not subtle. Cause was incorrect BIOS setting for serial ports. Unfortunately, the incident of the 19th, comprised of impossibly high means for perhaps a minute after a resumption following a break, means the data from 2222 may not be reliable in general, and should be excluded from primary analyses. Subsequent monitoring indicates the egg is functioning normally, so return to the reliable list as of July 1, 2001.
- Egg 1000, uncertain data after restart on February 27 2002, continuing to date. No obvious bad trials, but many missings. The uncertainty and the missings result from using a non-standard, untested system of eggsh and OS. It is the new MAC OS X. By trial and error, the host discovered the number of missings was reduced if the regtest program is running concurrently with eggsh. I recommend excluding the data for analyses that need guaranteed data, but otherwise have no reason to reject them.
- Egg 226 had a 0 in the data from 2002-10-29. Apparently a hardware crash generated at least one corrupt trial. Exclude the whole day.
- Egg 161 had bad data from about 02:15 to 08:30 on 2002-12-03. The egg was moved to new hardware about that time, but I don’t yet know what caused the problem. Exclude the whole day.
- Egg 1022 had bad data on 2003-02-24, for a short period, surrounded by no data. The computer had crashed and this was from a new installation, all high trials, ranging from 110 to 124, mean 118.121. The egg also had bad data on 2003-02-25. Exclude both days.
- Egg 2236 had bad data on 2003-03-11, reporting both a 0 and a 251, and a Variance of 55.722. Exclude 2236 the whole day.
- Egg 142 had bad data on 2003-05-21, reporting a 0 and a 184, after which it shut down. Exclude egg 142 for the whole day.
- Egg 227 had bad data on 2003-05-25, reporting a series of 255 values. It was being transferred to a new host and had bad data when it was
- Egg 2235 had a 0 on 2003-06-23. That whole day for 2235 should be excluded. The same thing on 2003-06024, so exclude that day also for 2235, and shut down the egg — as of June 26 00:10. Exclude also the 25th and 26th.
- Egg 102 had a 0 and a 250 in 85863 trials on 2003-08-06. That whole day for 102 should be excluded. The host reported his system was damaged by the heat wave and had repairs around this date.
- Egg 102 had a 253 in 42153 trials on 2003-08-13. That whole day for for 102 should be excluded. Host has been having trouble, see previous.
- Egg 2235 was opened again for data in August, and ran OK for some weeks, then had a 52 (6.8 sigma) on 2003-09-12, which indicates it still has intermittent errors. The whole day of 2003-09-12 should be excluded for that egg.
- Egg 110 had a 0, probably because of an unexpected shutdown, so exclude 2003-10-05 00:00:00,2003-10-05 23:59:59,110
- Egg 2235 had a 31 on 2003-10-07 and a 0 on 2003-10-08. Both days should be excluded for that egg. Also the 9th. For rigorous analyses, egg 2235 should be excluded for all days beginning 2003-06-23. Another 0 on 031013 and one on 031014. Exclude the 15th and 16th as well. Turned off on Oct 16 2003.
- Egg 115 had a 173 around 21:15, with no indication of the source. Exclude the whole day 2003-11-05 00:00:00,2003-11-05 23:59:59, egg 115.
- Egg 2236 had bad data beginning 2003-11-09, identified with low variance indicating failure of RNG or Serial port. Exclude 2236 from that date until further notice.
- Egg 115 had a 180 around 01:45 on Nov 16 2003, no indication of source. Exclude the whole day 2003-11-16 00:00:00,2003-11-16 23:59:59, egg 115.
Egg 100 had bad data generated during a shutdown on Dec 11 2003. Exclude the whole day 2003-12-11, 00:00:00 23:59:59, egg 100.
Please refer to the Official Error List for up to date information.
In addition, there are occasional trials that are clearly out of range, that apparently occur when something causes a computer shutdown. See the following sections.
Device Standard Deviation
Because the three REG/RNG devices are real-world machines, they cannot be
perfect theoretical devices. The logical XOR does guarantee that the mean will be unbiased, but it is possible for the standard deviation (SD) to be consequentially reduced if the logic corrects for a very large bias. Other factors can in principle affect the empirical SD, for example, bit-wise autocorrelation. The designs are excellent, so these potential distortions are very small, indeed too small to be detected in analyses of hours or days. However, with really large samples, using on the order of six months of data, several of the devices do show significant decreases or increases, and it is possible to see
SD signatures that distinguish the the three REG types. Mike Meyer has been carefully examining the actual behavior of the REG/RNG devices over long time periods, to see how much the real-world physical devices differ from their theoretical image. The most comprehensive version of his work is summarized in tables Comparing Theoretical and Empirical SD for GCP calculations. An earlier version of this effort looks at SD for individual eggs.
Out of Range Trials
Mike’s programs also checked for, and rejected all values of 0 and 200. In the period tested for the individual SD calculations, the results show four instances of rejection, and show that the zeros found have the same pattern surrounding them. The pattern for all 4 days is:
-a few other numbers
-then the generators produce no data for an unknown length of time
Here are the dates and times of the events:
Jun 22, 2000 21:43:16 egg id# 111 Dec 09, 2000 13:57:48 egg id# 115 Feb 27, 2001 01:33:04 egg id# 118 Mar 16, 2001 21:43:16 egg id# 106
The simplest procedure to deal with this infrequent but important misbehavior is to include a step in analysis that excludes all data with the value 145.
It is possible that other instances exist, although the general case is that the eggs either work according to design, or they break badly and show obviously defective data. We have discussed implementing some filtering to remove bad data according to objective criteria, but this has not been done. If you have ideas and skills which you would like to apply to this task, please let Roger Nelson know.
While most of the eggs run in complete synchrony to the second, using software to continually correct the time using timeservers or other precise sources, a few dialup eggs are apparently not running such programs, or cannot connect to timservers during their relatively short dialup times. Their data-streams are therefore not precisely aligned but are likely to be offset to varying degrees from correct UTC time. In most cases the difference is on the order of a few seconds, and will not make an important difference in analyses using large blocking units, although the offset will exert a conservative influence in any fine-scale analyses. The following list are probably off by a few seconds, and they should be excluded from analyses that strongly depend on synchronization to the second: Eggs 100, 103, 107, 108, 114, 1025, 1026.
In addition to actual errors, there are occasional differences from the normal course of events, such as the following instance of testing with a different random source than the one normally assigned to the node:
For EGG 2000, the period from UTC 16-Jan-01 13:53:59 until 17-Jan-01 04:24:08 was collected using Mindsong REG #139, rather than the normal Orion REG. There are short gaps in data before and after this period, so it should be clearly identifiable in the data. This substitution should not make any important difference in results; we have not detected any differences in the performance of the three different REG/RNG sources.
Until recently, that is. Mike Meyer examined individual egg parameters in data sequences of six months duration. He determined that while the mean is within specifications, the standard deviation is slightly different from theory. The effect is probably not important in most analyses. The magnitude is on the order 0.0001 on average. For best accuracy, the empirical standard deviation calculated over a large database (at least several days) should be used.