:The good news is there was no obvious stability issue; the tests completed
:successfully.
:But there was no real performance change either. blogbench results were
:624 / 176300, in line with previous ones.
I'll commit the change.
:> In anycase, I'm not sure it's even the bottleneck for the earlier tests
:> since blockbench didn't run long enough (with default options) to use
:> up 32G of ram.
:
:On the contrary, shouldn't performance be higher if the disks are not touched ?
:
:Looking at iostat history, newfs pushed almost 8000 iops to the raid volume.
:blogbench was much lower at first and the number of iops was almost divided
:by ten after some time.
Not necessarily, because the I/O's are going to be mostly asynchronous
writes and not synchronous reads.
Linear write iops is mostly irrelevant. That's just the platter write
bandwidth divided by the I/O size.
:I've attached a log of iostat /dev/da0 1
:
:Two activity periods are shown in the file:
:- newfs_hammer -L RAID_VOL /dev/da0
:- blogbench -d /mnt/blogbench
:
:--
:Francois Tigeot
blogbench --iterations=200 ... and you really need to include the
actual blogbench output. The final results are worthless.
: 0 17 16.00 5922 92.51 3 0 2 0 95
: 0 144 16.00 7439 116.24 3 0 2 0 94
: 0 6 16.00 7672 119.88 3 0 2 0 95
: 0 6 16.00 7316 114.32 3 0 3 0 94
Linear write activity. I wonder why it isn't
clustering the I/O's, though.
: 0 6 15.99 4762 74.38 5 0 95 0 0
: 0 5 15.99 2615 40.82 4 0 84 0 12
: 0 5 15.98 7683 119.86 3 0 73 0 24
: 0 6 15.64 1741 26.59 3 0 84 0 13
: 0 5 15.65 1310 20.01 3 0 70 0 27
: 0 5 15.81 1108 17.11 4 0 90 0 6
This isn't too good for a RAID volume. The TPS is there but it
should not be doing any mixed reading activity that early in the
blogbench test.
Try setting vfs.hammer.double_buffer=1.
: 0 5 15.75 252 3.87 2 0 80 0 18
: 0 11 15.80 240 3.71 1 0 84 0 15
: 0 5 15.74 310 4.76 1 0 86 0 13
: 0 5 15.79 231 3.56 1 0 83 0 16
:...
: 0 3 15.94 288 4.48 1 0 85 0 14
: 0 3 16.00 620 9.69 1 0 79 0 20
: 0 3 16.00 352 5.50 1 0 84 0 15
This doesn't look good either. It shouldn't degrade that much,
though it's a bit hard to tell with iostat because it doesn't print
the disk busy %.
-Matt
Matthew Dillon
<dil...@backplane.com>
