Andi Kleen's blog

Tilting at windmills and other endeavors

Archive for April, 2012

The weekend error anomaly

with 5 comments

I run which describes the standard way in Linux to handle machine check errors. Most of the hits of the website are just people typing a error log into a search engine ranks quite high on Linux machine check related terms.

The log files give me some indication how many errors are occurring on Linux systems in the field. Most of the errors are corrected memory errors on systems with ECC memory: a bit flipped, but the ECC code corrected it and and no actual data corruption occurred. (In this sense they are not actually errors, a more correct term would be “events”). Other errors like network errors or disk errors are not logged by mcelog.

I noticed is that there seem to be less memory errors on weekends. Normally the distribution of hits is fairly even over the week. But on Saturday and Sunday it drops into half.

It’s interesting to speculate why this this weekend anomaly happens.

ECC memory is normally only on server systems, which should be running 24h. In principle errors should be evenly distributed over the whole week.

Typing the error into google is no automated procedure. A human has to read the log files and do it manually.

If people are more likely to do this on work days one would expect that they would catch up on the errors from the weekend on Monday. So Monday should have more hits. But that’s not in the data: Monday is not different from other weekdays.

It’s also sticky for each system (or rather each human googling). Presumably the person will google the error only once no matter how many errors their system have and after that “cache” the knowledge what the error means. So the hits are more a indication of “first memory error in the career of a system administrator” (assuming the admin has perfect memory, which may be a bold assumption). But given a large enough supply of new sysadmins this should be still a reasonable indication of the true number of errors (at least on the systems likely to be handled by rookie administrators)

The hour distribution is more even, with 9-10 slightly higher. Not sure which time zone and what that means on the geographical distribution of errors and rookie admins.

One way to explain the weekend anomaly could be that that servers may be more busy on weekdays and they may have more errors when they are busy. Are these two assumptions true? I don’t know. It would be interesting to know if this shows up in other peoples large scale error collections too.

I wonder if it’s possible to detect solar flares in these logs. Need to find a good data source for them. Are there any other events generating lots of radiation that may affect servers? I hope there will never be a nuke or a super nova blast in the data.

Written by therapsid

April 20th, 2012 at 3:11 am

Posted in curiosities,kernel