13x longer execution times before remount

[Ian Lepore]

On Mon, 2012-05-14 at 21:23 +0200, Fredrik Sikén wrote:
> Hi,
> 
> I've an 8.2-Release running on an ARM with poor execution performance
> before a remount is performed.
> 
> The scenario is as follows
> 
> 0) I mount an UFS SSD disk
> # mount -o noclusterr -o noclusterw /dev/ad0s1 /mnt/
> 
> 1) I copy an executable to an UFS mounted SSD disk and measure its
> time of execution from start to stop
> 
> 2) I remount the system (umount/mount)
> 
> 3) I perform the same steps as in 1) but now with a performance increase of 13x
> 
> Other tests performed:
> - Tested on a UFS USB stick with the same behaviour
> - Tested to write to an NFS mounted disk and this gives the expected
> behaviour - the same time before and after a remount.
> - Running the same version on a x86 gives the expected behaviour - the
> same time before and after a remount.
> 
> Results after file is copied but before a remount
> # time ./test 10000 1000000
> size = 10000, no = 1000000
> 133.000u 0.000s 2:13.53 99.8%   3282+18575k 0+0io 0pf+0w
> 
> Results after a remount
> # time ./test 10000 1000000
> size = 10000, no = 1000000
> 7.000u 0.000s 0:07.82 99.8%     3286+18595k 0+0io 0pf+0w
> 
> The test application source looks like this
> #include <string.h>
> #include <stdlib.h>
> #include <stdio.h>
> 
> int main(int argc, char **argv) {
> 
>   int size = atoi(argv[1]);
>   int no = atoi(argv[2]);
>   printf("size = %d, no = %d\n", size, no);
> 
>   {
>     int i;
> 
>     for(i = 0 ; i<no ; i++) {
>       char *p = malloc(size);
>       memset(p, 0, size);
>       free(p);
>     }
>   }
> }
> 
> Thanks,
> Fredrik

More info on this problem and a potential workaround (not ideal, but
we've been living with it) can be found here:

http://lists.freebsd.org/pipermail/freebsd-arm/2012-January/003288.html
http://lists.freebsd.org/pipermail/freebsd-arm/2012-February/003296.html

- Ian