This document briefly describes the
performance tests done during the development of 64 bits extents.
The development is done in three phases :
Phase 1: convert on-disk structures to architecture independent data
using little endian values (not specific to 64 bits in fact).
Phase 2: use 64 bits values for addressing physical blocks.
Phase 3: use 64 bits values for addressing logical blocks.
At each phases, the test ran with and without the new code to see
the performance impact of the new development.
Here after, the tests
done at each phase and the results are described.
The tests are basic high level tests using
cp and
rm. The system time
consumed by both commands is
measured with
time.
All development and tests are done on kernel version 2.6.11.7 with
extents version 0.4.
Patches:
The patch for the 1st phase is
ext3-extents-little_endian.patch.
The patch for the 2nd phase is
ext3-extents-0.4-le-physical-block64-2.6.11.7.patch.
The last patch for the 3rd phase is
ext3-extents-0.4-logical-block64-2.6.11.7.patch.
These patches must be applied in that order on top of Alex Tomas patch
ext3-extents-2.6.11-0.4.patch.
Phase 1: architecture independence of on-disk structure
As the format used for on-disk data is little endian, the tests have
been done on power machine, which natively works with big endian values.
The test is done on a PL250 (dual Power 5 with 4 GB of RAM).
It uses a loopback file, which is filled and emptied with
/usr (about 2GB of data) as
follow:
dd if=/dev/zero of=~/loop.img bs=1k count=2000000 2>&1 1>/dev/null
losetup /dev/loop0 ~/loop.img
mkfs.ext3 /dev/loop0 2>&1 1>/dev/null
mount -o extents /dev/loop0 /mnt
TIME1=`/usr/bin/time -f "%U %S %e" cp -r /usr /mnt 2>&1`
umount /dev/loop0
mount -o extents /dev/loop0 /mnt
TIME2=`/usr/bin/time -f "%U %S %e" rm -rf /mnt/usr 2>&1`
umount /dev/loop0
losetup -d /dev/loop0
echo "cp: $TIME1 rm: $TIME2"
|
The test ran 100 times with and without code modification. The
results before the development are
here
and the one after are
here.
At each
line, the values indicate the user-, the system- and the real time
consumed for the command.
On average, the results are :
|
cp
|
rm
|
Before
|
After
|
Difference
|
Before
|
After
|
Difference
|
User time
|
0.46 |
0.45 |
|
0.05 |
0.05 |
|
System time
|
17.98 |
18.10 |
+0.67%
|
3.47 |
3.49 |
+0.58%
|
Elapse time
|
206.60 |
207.30 |
|
3.54 |
3.55 |
|
Phase 2: use 64 bits values for addressing physical blocks
The test was run on a dual Xeon 2.8GHz with 2GB of RAM.
The system was installed on the first disk. The test consists
of repeating several copy/remove on the tested filesystem installed on
a dedicated
disk/partition.
The script used is quite the same as before:
mkfs.ext3 /dev/sdc1 2>&1 1>/dev/null
mount -o extents /dev/sdc1 /mnt
TIME1=`/usr/bin/time -f "%U %S %e" cp -r /usr /home /boot /lib /bin /etc /sbin /var /mnt`
umount /dev/sdc1
mount -o extents /dev/sdc1 /mnt
TIME2=`/usr/bin/time -f "%U %S %e" rm -rf /mnt/usr /mnt/home /mnt/boot /mnt/lib /mnt/bin /mnt/etc /mnt/sbin /mnt/var`
umount /dev/sdc1
echo "cp: $TIME1 rm: $TIME2"
|
The previous script has been run 400 times on the machine without other
special activities. The amount of data copied each time is about 16 GB.
The results before are
here
and the one after are
here
(the format of the results is as in previous section).
On average, the results are:
|
cp
|
rm
|
Before
|
After
|
Difference
|
Before
|
After
|
Difference
|
User time
|
0.74 |
0.75
|
|
0.10
|
0.09
|
|
System time
|
14.32
|
14.30
|
- 0.14%
|
2.55
|
2.55
|
+0.00%
|
Elapse time
|
126.70
|
126.65
|
|
13.86
|
13.88
|
|
Phase 3: use 64 bits values for addressing logical blocks.
The test was run on a dual Xeon 3.0GHz with 4GB of RAM.
As before, the system was installed on the first disk and the tests
were ran on the third one.
The test done is exactly the same. The previous script has been
repeated 330 times in both cases (with and without development) and the
amount of copied data each time is about 4 GB.
The results are
here
(before)
and
here (after).
On average, it gives the following result:
|
cp
|
rm
|
Before
|
After
|
Difference
|
Before
|
After
|
Difference
|
User time
|
1.19
|
1.21
|
|
0.17
|
0.17
|
|
System time
|
24.54
|
24.64
|
+0.40%
|
5.04
|
5.03
|
-0.20%
|
Elapse time
|
298.74
|
298.39
|
|
40.58
|
40.61
|
|