Re: About ceph_clock_now()

["Adam C. Emerson" <aemerson@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2922 bytes --]

On 12/01/2016, Erwan Velu wrote:
> Hi folks,

Hello!

> When reading the code I noticed that ceph_clock_now() is mostly used to compute a local difference of time.
> I do see you are using "clock_gettime(CLOCK_REALTIME, &tp)" but was wondering if CLOCK_MONOTONIC_RAW isn't more accurate.
> 
> I mean if the local system time is readujsted by NTP (or similar) some computation could be wrong leading to some possible mis-interpretation or wrong scheduling.

I have done some work starting in that direction currently in master. In
common/ceph_time.h I define coarse and fine versions of the real clock
and monotonic clock and I've switched the osdc subsystem over to use them.

We use the std::chrono C++ time library, but we define our own clocks
(so we can pick the right one for performance/resolution tradeoffs) and an
in-memory timespan representation (so not every function taking a duration has
to be a template.)

The real_clock and mono_clock classes correspond to CLOCK_REALTIME and CLOCK_MONOTONIC, and coarse_real_clock and coarse_mono_clock correspond to CLOCK_REALTIME_COARSE and CLOCK_MONOTONIC_COARSE. If you require millisecond resolution but nothing finer, you should prefer the coarse variants.

The std::chrono library lets us integrate with other C++ libraries (and use the
standard concurrency features) and has the advantage of stronger typing; we
can't use a duration where a time point is required or subtract a monotonic
time point from a real time point.

As I understand it, CLOCK_MONOTONIC_RAW is generally the wrong choice.
The raw hardware clock is prone to drift, and NTP synchronization (or other
corrective) makes the elapsed time measurements more accurate.

However, I am not a Time Lord. If someone knows better, please correct me here.

One concern with monotonic clocks, even for elapsed time, is that their epoch is
unspecified. You cannot send a monotonic time over the network to another
machine nor can you save it to a file and re-read it later. Anything to be
serialized must be a real time.

This is unfortunate. If I were a Time Lord I'd travel back in time and make the
POSIX committee adopt CLOCK_UTC and CLOCK_TAI so we could have serializable
monotonic time values.

> I could work on a PR is you find this interesting.

I find it interesting. If you want to change more subsystems to use ceph_time
and use monotonic and coarse clocks in appropriate places I think that would be
wonderful. It might be worthwhile to focus on the places where we use double to
represent a time value.

Please read through ceph_time.h and osdc to get a feel for it. If you think it
could use improvement, I'd be happy to hear your ideas.

Thank you.

-- 
Senior Software Engineer           Red Hat Storage, Ann Arbor, MI, US
IRC: Aemerson@{RedHat, OFTC, Freenode}
0x80F7544B90EDBFB9 E707 86BA 0C1B 62CC 152C  7C12 80F7 544B 90ED BFB9

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 603 bytes --]