On Mon, 9 Mar 2015 17:08:58 +0100
Michael Holzheu <holzheu at linux.vnet.ibm.com> wrote:
> Hello Petr,
>
[snip]
> As a conclusion, we could think of mapping larger chunks
> also for the fragmented case of -d 31 to reduce the amount
> of mmap/munmap calls.
>
FYI: I did some more tests and I am no longer sure if the above
conclusion was correct.
A simple "copy" program that reads or maps/unmaps every page
from /proc/vmcore and then writes it to /dev/null is faster
with mmap()/munmap() than with using read():
read:
-----
# time ./copy /dev/null read
real 0m1.072s
user 0m0.010s
sys 0m1.054s
# perf stat -e syscalls:sys_enter_old_mmap,syscalls:sys_enter_munmap,syscalls:sys_enter_read ./copy /dev/null read
8 syscalls:sys_enter_old_mmap
1 syscalls:sys_enter_munmap
458753 syscalls:sys_enter_read
1.405457536 seconds time elapsed
mmap:
-----
# time ./copy /dev/null mmap
real 0m0.947s
user 0m0.314s
sys 0m0.631s
# perf stat -e syscalls:sys_enter_old_mmap,syscalls:sys_enter_munmap,syscalls:sys_enter_read ./copy /dev/null mmap
458760 syscalls:sys_enter_old_mmap
458753 syscalls:sys_enter_munmap
1 syscalls:sys_enter_read
1.175956735 seconds time elapsed
Regards,
Michael
