On Sat, Jun 10, 2017 at 9:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> We already do call getrusage(). The point of that comment is that the
> contents of the resulting struct rusage are not very well standardized.
> POSIX says only
>
> The <sys/resource.h> header defines the rusage structure that includes
> at least the following members:
>
> struct timeval ru_utime user time used
> struct timeval ru_stime system time used
>
> (seems the same in 1997 and 2008 text). So the existing code has already
> got its neck stuck way out in terms of portability. Maybe you could push
> it further, but I bet you'll find that the situation isn't any better than
> it was at the time that comment was written.
>
> It's entirely possible that we could simplify the code some, because
> I suspect that Windows is the only remaining platform that doesn't
> HAVE_GETRUSAGE. But that in itself doesn't mean we can use any more
> fields than we do now.
It might be worth adding platform-specific code for common platforms.
I checked macOS X 10.10.5 (Yosemite) and a Linux system (hydra, old
Fedora release) and they seem to list identical fields for struct
rusage, which is good as far as it goes, but Linux documents a bunch
of the interesting fields (ru_ixrss, ru_idrss, ru_isrss, ru_nswap,
ru_msgsnd, ru_msgrcv, ru_nsignals) as unused, and apparently returns
ru_maxrss in kilobytes where macOS reports it in bytes. It seems like
it would be a good idea to install code specific to Linux that
displays all and only those values that are meaningful on Linux, and
(less importantly) similarly for macOS. Linux is such a common
platform that reporting bogus zero values and omitting other fields
that are actually meaningful does not seem like a very good plan.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
