> > And here are the dd results for 16GB RAM, i.e. 4,000,000 8K blocks:
>
> So, if we divide 32,000 MB by the real time, we get:
> /data (data):
> 89 MB/s write
> 38 MB/s read
... snip ...
> The read speed on your /data volume is awful to the point where you should
> consider it broken and find a fix. A quick comparison: the same number on a
> 16 drive internal SATA array with 7200 RPM disks gets 950 MB/s read, about
> 25 times faster for about 1/4 the price.
I managed to get approval to shut down the Oracle instance and reran
the dd's on the SAN (/data) and came up with about 60MB/s write (I had
missed the 'sync' in the previous runs) and about 58 MB/s read, still
no comparison on your SATA arrary. Any recommendations on what to
look at to find a fix? One thing which I never mentioned was that I
am using ext3 mounted with noatime,data=writeback.
An interesting note (at least to me) is the inverse relationship
between free memory and bo when writing with dd, i.e:
$ vmstat 5
r b swpd free buff cache si so bi bo in cs us sy id wa
0 3 244664 320688 23588 15383120 0 0 0 28 1145 197 0 1 74 25
2 6 244664 349488 22276 15204980 0 0 0 24 1137 188 0 1 75 25
2 6 244664 28264 23024 15526552 0 0 0 65102 1152 335 0 12 60 28
2 4 244664 28968 23588 15383120 0 0 1 384386 1134 372 0 19 34 47
1 5 244664 28840 23768 15215728 0 0 1 438482 1144 494 0 24 33 43
0 5 247256 41320 20144 15212788 0 524 0 57062 1142 388 0 6 43 51
1 6 247256 29096 19588 15226788 0 0 5 60999 1140 391 0 15 42 43
Is this because of the kernel attempting to cache the file in memory?