Thanks for your bare results.
Just try the following test:
1 - Create 2 files of about the size of your system main memory.
My config is 32MB
total 65806
-rwxr-xr-x 1 root root 33554432 Jul 8 21:02 1*
-rwxr-xr-x 1 root root 33554432 Jul 8 21:02 2*
drwxr-xr-x 2 root root 12288 Jun 21 22:57 lost+found/
2 - Run the following script 2 times with tagged queue disabled and 2
times with tagged queue enabled (4 tags min).
time cat 1 >/dev/null &
time cat 2 >/dev/null &
Here are the results I got with may IBM S12 / 512KB cache.
Tagged command queuing disabled:
--------------------------------
0.02user 1.41system 0:22.64elapsed 6%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8243major+15minor)pagefaults 0swaps
0.00user 1.39system 0:22.79elapsed 6%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8242major+15minor)pagefaults 0swaps
0.01user 1.65system 0:22.39elapsed 7%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8238major+15minor)pagefaults 0swaps
0.00user 1.44system 0:22.43elapsed 6%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8242major+15minor)pagefaults 0swaps
Tagged command queuing enabled:
-------------------------------
0.01user 1.50system 0:13.36elapsed 11%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8246major+15minor)pagefaults 0swaps
0.02user 1.57system 0:18.16elapsed 8%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8241major+15minor)pagefaults 0swaps
0.00user 1.65system 0:12.91elapsed 12%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8248major+15minor)pagefaults 0swaps
0.02user 1.54system 0:17.79elapsed 8%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (8243major+15minor)pagefaults 0swaps
If you donnot gain advantage of tagged queue, you can try to decrease
asynchronous read-ahead max size.
Just edit mm/filemap.c and apply the following change:
-#if 0 /* small readahead */
+#if 1 /* small readahead */
#define MAX_READAHEAD (PAGE_SIZE*7)
#define MIN_READAHEAD (PAGE_SIZE*2)
#else
#define MAX_READAHEAD (PAGE_SIZE*18)
#define MIN_READAHEAD (PAGE_SIZE*3)
#endif
If bonnie would include this test, we would get:
1.5 MB/sec with tagged queue disabled
2.2 MB/sec with tagged queue enabled
IMHO, this benchmark for tagged command queuing is intelligent enough for
Bonnie.
With realist tests for tagged command queuing, my IBM S12 does not
gain advantage of this feature (about 2% to 4%).
However, with some "a la Bonnie" test (see above), performances are
(2.2-1.5)/1.5 -> 46% better with tagged queue enabled.
I invite all bonnie benchmen to try the above test with different drives,
adapters and drivers.
On Sun, 7 Jul 1996, Robert L Krawitz wrote:
> My experience with the BSD driver in 2.0.3: large stream writes are
> slightly faster than with Drew's (6.0 vs. 5.6 MB/sec with my
> configuration). Stream reads are somewhat slower (5.2 vs. 6.4
> MB/sec). In both cases, the BSD driver used more CPU time (about 7
Can you send to me the conditions of the stream reads tests, please.
> seconds for 128 MB vs. 6 with Drew's driver). The random seek portion
> of Bonnie may have been to the BSD driver's advantage, but not by much
> (110-120 ops/sec vs. 95-105 with Drew's driver).
>
> As for the real test -- startup of X, including Netscape and emacs --
> neither driver had an obvious advantage.
>
> --
> Robert Krawitz <rlk@tiac.net> http://www.tiac.net/users/rlk/
Gerard.