svn commit: r184471 - in head/bin: cat cp

[Giorgos Keramidas]

On Thu, 30 Oct 2008 14:05:58 +0000 (UTC), Ivan Voras <ivoras at freebsd.org> wrote:
> Author: ivoras
> Date: Thu Oct 30 14:05:57 2008
> New Revision: 184471
> URL: http://svn.freebsd.org/changeset/base/184471
>
> Log:
>   Teach cat(1) and cp(1) to use a larger buffer if enough memory is present
>   in the system. A simple heuristics is used to detect what is "enough"
>   memory: if number of physmem pages is greater than 32k (equalling 128 MB
>   on machines with 4 kB pages).
>
>   Typical immediate result of these changes is reduction in context switches
>   and the goal is to increase efficiency by using large buffers:
>   before: /usr/bin/time -hlp cat file1 > file2
>   ...
>          163  voluntary context switches
>        11194  involuntary context switches
>   after: /usr/bin/time -hlp ./cat file1 > file2
>   ...
>          417  voluntary context switches
>          272  involuntary context switches

Nice :)

I didn't notice this in -hackers, but it may be worth letting the buffer
size grow above 1 MB when we have more than, say, 512 MB of memory.  By
running dd(1) with buffer sizes of 1, 2, 4, 6, 8, 10, 12, 14, and 16 MB
there are larger buffer sizes that reduce involuntary context switches
even more.

My laptop (with more than 2 GB or memory), with powerd disabled, running
in single user mode, and copying a file of 200 MB using the script:

,-----------------------------------------------------------------------
| #!/bin/sh
|
| mblist="1 2 4 6 8 10 12 14 16"
|
| for bufsize in ${mblist} ; do
|         cmd="dd if=testfile bs=${bufsize}m | dd of=testfile.out bs=${bufsize}m"
|         logname=`printf '%02d.dd-times' "${bufsize}"`
|         true > "${logname}"
|         for cnt in `jot 30 1 30` ; do
|                 echo "Overwriting testfile"
|                 cat datafile > testfile
|                 printf '%2d MB buffer, run %2d, logfile=%s\n' \
|                     "${bufsize}" "${cnt}" "${logname}"
|                 /usr/bin/time -l sh -c "${cmd}" 2>> "${logname}"
|                 tail -1 "${logname}"
|         done
| done
`-----------------------------------------------------------------------

got the following results for sizes < 10 MB:

: tcsh% ministat -w 78 0?
: x 01
: + 02
: * 04
: % 06
: # 08
: +------------------------------------------------------------------------------+
: |     #               % #% *   *      +         x                              |
: |    ##    %#*#* %%%  # O%*O% ***+*  *+ + + x   x   x    x xx   x              |
: |## %### %%O#O#O##OOO#OOOOOOO *OO**@*O**++*xx x x xxx  x * xx   x x   x  x    x|
: |      |___|____AM|MA____||A|___A_M_|__| |__________MA__________|              |
: +------------------------------------------------------------------------------+
:     N           Min           Max        Median           Avg        Stddev
: x  30          1383          1653          1493        1496.1     69.360801
: +  30          1286          1434          1381     1371.0667     44.393952
: Difference at 95.0% confidence
:         -125.033 +/- 30.1005
:         -8.35728% +/- 2.01193%
:         (Student's t, pooled s = 58.2312)
: *  30          1242          1523          1342     1339.3667     56.059379
: Difference at 95.0% confidence
:         -156.733 +/- 32.5975
:         -10.4761% +/- 2.17883%
:         (Student's t, pooled s = 63.0618)
: %  30          1199          1401          1292     1293.1667     50.603996
: Difference at 95.0% confidence
:         -202.933 +/- 31.3824
:         -13.5642% +/- 2.09762%
:         (Student's t, pooled s = 60.7111)
: #  30          1179          1402          1275     1273.9667     57.161889
: Difference at 95.0% confidence
:         -222.133 +/- 32.8523
:         -14.8475% +/- 2.19586%
:         (Student's t, pooled s = 63.5547)

Buffer sizes even larger didn't yield as good results as 6 MB or 8 MB,
but there seems to be an improvement in 16 MB too:

: tcsh% ministat -w 78 01 1?
: x 01
: + 10
: * 12
: % 14
: # 16
: +------------------------------------------------------------------------------+
: |                 #    #         %    % %                                      |
: |          # # #  ##*  ##  O#+% OO*   %+%*  +  %*+ x*x x   x xx  x             |
: |# ##   #  #OO*#O OO* *OO%OOOOOOOO*OOOOOOO**@ xOO*%x*#%* * x xx  x x   x  x   x|
: |         |________A|_____||_|_AM___AAMM___|_||________A_________|             |
: +------------------------------------------------------------------------------+
:     N           Min           Max        Median           Avg        Stddev
: x  30          1383          1653          1493        1496.1     69.360801
: +  30          1200          1511          1383     1363.5667     69.974716
: Difference at 95.0% confidence
:         -132.533 +/- 36.0126
:         -8.85859% +/- 2.4071%
:         (Student's t, pooled s = 69.6684)
: *  30          1208          1496          1336     1333.1667     79.330858
: Difference at 95.0% confidence
:         -162.933 +/- 38.5167
:         -10.8905% +/- 2.57447%
:         (Student's t, pooled s = 74.5128)
: %  30          1208          1488          1377     1374.1333     59.381138
: Difference at 95.0% confidence
:         -121.967 +/- 33.3741
:         -8.15231% +/- 2.23074%
:         (Student's t, pooled s = 64.5641)
: #  30          1124          1363          1246     1245.7667     59.806979
: Difference at 95.0% confidence
:         -250.333 +/- 33.4755
:         -16.7324% +/- 2.23752%
:         (Student's t, pooled s = 64.7603)

It looks like it's probably worth pushing the limit up to 8 MB when we
have lots of memory :)