Comments on: What time 18446744073709550.000 means

By: peter

peter — Wed, 20 May 2009 16:45:45 +0000

Vadmin – do you know why Google patches do not use monotonic clock in their patches for wall clock timing ?

By: peter

peter — Wed, 20 May 2009 16:45:09 +0000

Tim,

I think math wise you’re right. However on the practical measures dealing with the huge values (not small negative values) is quite inconvenient. We also mainly use this for things like query execution time etc where the gain of accuracy by using negative values is not worth loss of convenience.

By: foo

foo — Tue, 19 May 2009 22:02:00 +0000

peter,

a monotonic clock won’t go backwards and it probably provides equal or better performance when compared to the other clocks supported by clock_gettime.

By: http://blog.timbunce.org/

http://blog.timbunce.org/ — Tue, 19 May 2009 21:23:44 +0000

Vadim. I believe the clocks are synchronized, just not as frequently as we’d like so they drift slightly. I had a link that, as I recall, explained that, but it doesn’t work now (http://webnews.giga.net.tw/article/mailing.freebsd.performance/710).

Re your analogy, imagine you have two watches (2 cpus) and each time you measure an interval you pick a watch at random (thread switched to different cpu). Over thousands of measurements you’ll get a more accurate total if you don’t discard negative durations.

By: Vadim

Vadim — Tue, 19 May 2009 17:28:28 +0000

TimBunce,

It is two different timers. They are not synchronized. If we get negative time from two different clocks I do not see why we should add this time to total time, it has no good information.
It’s like you measure time for runner and measure start time by your watches and measure finish time by your friend’s watches, and difference between both watches is unknown.
It may happen your friend’s watches is slower by XYZ sec. And by your watches runner started at 10:54 and by your friend’s watches he finished at 10:37.
By subtracting you get -0:17, but does it really have any sense ?

By: Shlomi Noach

Shlomi Noach — Tue, 19 May 2009 17:24:57 +0000

@Guillaume – thanks!

By: http://blog.timbunce.org/

http://blog.timbunce.org/ — Tue, 19 May 2009 16:51:27 +0000

Barron, I was referring to the underlying negative values that could arise from the clock_gettime() calls in the original article. The fact they’re presented as a huge positive my mysqld is obviously a bug.

By: Guillaume Lefranc

Guillaume Lefranc — Tue, 19 May 2009 12:58:27 +0000

@Shlomi, regarding negative times in slow query logs I filed this bug http://bugs.mysql.com/bug.php?id=35396, a patch was committed and should solve the problem. Nevertheless the problem was more about the NTP source that was not synchronizing correctly (ask your sysadmin not to run NTP in a virtual machine, quite a bad idea) and my servers having from -10/+10 sec. time readjust.

By: Baron Schwartz

Baron Schwartz — Tue, 19 May 2009 12:55:37 +0000

If we allow negative times to accumulate naturally, it will skew things horribly. These times go negative because they’re very small to begin with; it would take many of them to add up enough to cancel out a single negative-gone-huge-positive.

Maatkit has handled these for a long time in mk-query-digest by setting them to zero during log aggregation. I think this is reasonable because if it goes negative, it means it wasn’t measurable anyway.

By: http://blog.timbunce.org/

http://blog.timbunce.org/ — Tue, 19 May 2009 09:09:54 +0000

In general you should allow negative times to be accumulated naturally as they’ll tend to balance out over the longer term. Treating them as zero at the time they’re measured would be wrong. Treating them as zero in reports, after accumulation, is less wrong but it’s still better to educate the users to the harsh realities of clocks on SMP/virtual systems.

I’d strongly suggest you allow the clock id passed to clock_gettime() to be configurable. Then the user can make their own choice about the accuracy vs performance trade-off that suits their needs best.

It seems to be difficult to find good detailed information about the semantics of the various CLOCK_*’s that clock_gettime() supports on various systems. If anyone has better links I’d be grateful for them.