Hi,
I ran a set of benchmarks over different combinations
of ext3 journaling on mirrored disks. I got some
non-intuitive results. Could someone help me reason this ?
Benchmarks
----------
I ran concurrently 4 processes, each process having a
working set of 10MB.
a) Process 0 creates 160 files each of size 64KB.
b) Process 1 reads 160 files each of size 64KB.
c) Process 2 creates 2560 dirs. Each dir is an
empty dir but has a data block allocated to it.
So, each dir is of size 4KB
d) Process 3 stats 2560 files each of size 4KB.
All the processes executes a sync() call before exiting
and flushes all the data to disk.
I created a RAID 0 (stripping) over two RAIDS each of which
is a RAID 1 (totally there are 4 disks).
I measured the performance of these benchmarks on three
different setups (results are averaged over 20 runs).
I) Mirroring with data journaling (data=journal)
II) Mirroring with just metadata journaling (data=ordered)
III) Mirroring with no journaling.
Results
-------
Each column gives the time taken in seconds.
data=journal data=ordered NO journaling
CREATE: 3.47096 5.46494 1.745721
READ: 3.11347 5.09944 1.80708
MKDIR: 3.38010 5.48238 1.7545
STAT: 2.96075 4.27023 1.98247
In the results, mirroring with no journaling performs better
than the other two (as expected). However, the results look such
that "mirroring with data journaling" performs better than
"mirroring with metadata journaling". I don't know why
this is happening.
Any help is appreciated.
Thanks.
Vijayan
