Profiling git's test suite, Linus noted [1] that a disproportionately
large amount of time was spent reading /proc/meminfo. This is done by
the glibc functions get_phys_pages and get_avphys_pages, but they only
need the MemTotal and MemFree fields, respectively. That same
information can be obtained with a single syscall, sysinfo, instead of
six: open, fstat, mmap, read, close, munmap. While sysinfo also
provides more than necessary, it does a lot less work than what the
kernel needs to do to provide the entire /proc/meminfo. Both strace -T
and in-app microbenchmarks shows that the sysinfo() approach is
roughly an order of magnitude faster.
sysinfo() is much older than what glibc currently requires, so I don't
think there's any reason to keep the old parsing code. Moreover, this
makes get_[av]phys_pages work even in the absence of /proc.
Linus noted that something as simple as 'bash -c "echo"' would trigger
the reading of /proc/meminfo, but gdb says that many more applications
than just bash are affected:
Starting program: /bin/bash "-c" "echo"
Breakpoint 1, __get_phys_pages () at ../sysdeps/unix/sysv/linux/getsysstats.c:283
283 ../sysdeps/unix/sysv/linux/getsysstats.c: No such file or directory.
(gdb) bt
So it seems that any application that uses qsort on a moderately sized
array will incur this cost (once), which is obviously proportionately
more expensive for lots of short-lived processes (such as the git test
suite).
[1] http://thread.gmane.org/gmane.linux.kernel/2019285
Signed-off-by: Rasmus Villemoes <rv@rasmusvillemoes.dk>
* sysdeps/unix/sysv/linux/getsysstats.c (__get_phys_pages):
Use sysinfo system call instead of parsing /proc/meminfo.
* sysdeps/unix/sysv/linux/getsysstats.c (__get_avphys_pages):
Likewise.
