Here are the latest results for my elv_test elevator benchmark.
This benchmark gives three numbers, a baseline figure for writing
sequentially and write and read results for writing and reading in a
pattern designed to give the elevator a hard time.
The source code is available at
http://tltsu.anu.edu.au/~robert/elv_test.c.
I was hoping to see some improvements in 2.4.0-test9 based on recent
comments in the linux-kernel mailing list. But I havent seen any
improvement from 2.4.0-test9pre2.
The fundamental problem appears to be that disk writes are being
performed in a inefficient order causing the disk to seek far too often.
For comparisom, Ive done tests with kernel versions going all the way
back to 2.0.39 and discovered a few interesting points.
I will summarise first, full results are at the end.
Kernel 2.0.39 had a relatively slow read elevator (3 Mb/s read).
Kernel 2.2.14 improved on this (6 Mb/s read), but recent 2.2 kernels
(2.2.17+) have a problem with the read performance (back to 3Mb/sec).
Interestingly this doesnt seem to be an elevator problem because it
shows up worse in tiotest (1 Mb/sec read).
Kernel 2.4.0-test1 had quite good read and write elevator performance.
The write elevator dropped in 2.4.0-test3 to 2.5-3 Mb/sec.
And dropped again in 2.4.0-test9pre2 to 1 Mb/sec.
Annoyingly this elevator performance problem doesnt seem to be the only
issue that is effecting the netatalk benchmark that I started this
investigation with. The only kernel I see reasonable netatalk
performance from was 2.4.0-test1-ac22.
So there are several kernels that show reasonable performance in
elv_test which still perform badly in netatalk.
Results
=======
Kernel 2.4.0-test9
[robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 8.24 megs/sec
finished writing 240 megs written at 0.84 megs per sec
finished reading, 240 megs read at 5.34 megs/sec
Kernel 2.4.0-test9pre8 + 2.4.0-test9pre8-vm
[robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 8.44 megs/sec
finished writing 240 megs written at 0.86 megs per sec
finished reading, 240 megs read at 5.56 megs/sec
Kernel 2.4.0-test9pre2
robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 10.76 megs/sec
finished writing 240 megs written at 0.99 megs per sec
finished reading, 240 megs read at 5.58 megs/sec
Kernel 2.4.0-test7
[robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 8.70 megs/sec
finished writing 240 megs written at 2.41 megs per sec
finished reading, 240 megs read at 4.78 megs/sec
Kernel 2.4.0-test3
[robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 8.02 megs/sec
finished writing 240 megs written at 3.39 megs per sec
finished reading, 240 megs read at 5.02 megs/sec
Kernel 2.4.0-test1
[robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 7.97 megs/sec
finished writing 240 megs written at 6.90 megs per sec
finished reading, 240 megs read at 6.81 megs/sec
Kernel 2.2.18pre15
robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 9.56 megs/sec
finished writing 240 megs written at 8.57 megs per sec
finished reading, 240 megs read at 3.40 megs/sec
Kernel 2.2.17
[robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 8.33 megs/sec
finished writing 240 megs written at 7.65 megs per sec
finished reading, 240 megs read at 3.16 megs/sec
Kernel 2.2.14
robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 9.92 megs/sec
finished writing 240 megs written at 8.87 megs per sec
finished reading, 240 megs read at 6.86 megs/sec
Kernel 2.0.39
robert@testmac25 src]$ ./elv_test 8 30
files created, 240 megs written at 8.35 megs/sec
finished writing 240 megs written at 7.43 megs per sec
finished reading, 240 megs read at 3.88 megs/sec
-- Robert Cohen Unix Support, TLTSU Australian National University - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Oct 07 2000 - 21:00:13 EST