Hello,
In
http://groups.google.de/groups?selm=20020416165358.E29747%40dualathlon.random&output=gplain
Andrea Arcangeli forwarded my previous mail to the Linux kernel
mailing list. I have reported a serious IO problem of 2.4.18 and
2.4.19-pre5 where writing to DVD-RAM blocked reading from HDD almost
totally.
Here is the update of the benchmark for 2.5.40. It has _much_ improved
but is not yet where it could be. The column "factor 2.5.40" contains
the factor 2.5.40 is slower than it could theoretically be (so 1 would
be ideal and, in my opinion, achievable). The factor for 2.4.19-pre5
is given for comparison.
nr bench read write time expected factor factor same IDE
2.5.40 2.5.40 2.5.40 2.4.19-pre5 channel
1 dd 30GB HDD DVD-RAM 187 86 2.2 8.2 *
2 dd 120GB HDD DVD-RAM 57 31 1.8 14
3 dd 30GB HDD ZIP 87 86 1.0 4.0
4 dd 120GB HDD ZIP 75 31 2.4 7.8 *
5 dd 30GB HDD 120GB HDD 121 86 1.4 1.5
6 dd 120GB HDD 30GB HDD 65 31 2.1 2.2
7 cp 30GB HDD 120GB HDD 94 86 1.1 1.3
8 cp 120GB HDD 30GB HDD 69 62 1.1 1.3
The column "time 2.4.40" contains the execution time for the
benchmark, and "expected 2.4.40" contains the execution time I would
have expected for an ideal kernel.
The "dd" benchmark reads (with dd) a 1 GB file from the device in
column "read" while writing (also with dd) to the device in column
"write".
I have four IDE devices installed in that system:
hda: 30 GB IDE HDD 5400 rpm
hdb: 9.4 GB ATAPI DVD-RAM
hdc: 120 GB IDE HDD 7200 rpm
hdd: 100 MB ATAPI ZIP
The column "same IDE channel" indicates whether the device written to
an read from are on the same IDE controller.
Theoretically, the two devices (e.g. the 120 GB HDD and the DVD-RAM in
nr. 2) are independent (at least when they are on a different IDE
controller), so the read should perform as good as without the
additional writing activity (31 seconds, as given in the column
"expected"). But in reality the read takes 1.8 times as long. With
2.4.50, there could be performance gains up to a factor of 2.4 (test
nr. 4). But that's _much_ better than 2.4.19-pre5 where performance
gains of up to a factor of 14 (test nr. 2) were possible.
The "cp" benchmark simply copies a 1 GB file from "read" to "write"
and performs almost as good (factor 1.1) as it could (factor 1) with
2.4.50. Good work!
Here are the numbers for 2.4.18 and 2.4.19-pre5 for comparison.
nr bench read write time time expected factor
2.4.18 2.4.19-pre5 2.4.19-pre5 2.4.19-pre5
1 dd 30GB HDD DVD-RAM 278 490 60 8.2
2 dd 120GB HDD DVD-RAM 197 438 32 14
3 dd 30GB HDD ZIP 158 239 60 4.0
4 dd 120GB HDD ZIP 142 249 32 7.8
5 dd 30GB HDD 120GB HDD 87 89 60 1.5
6 dd 120GB HDD 30GB HDD 66 69 32 2.2
7 cp 30GB HDD 120GB HDD 97 77 60 1.3
8 cp 120GB HDD 30GB HDD 78 65 50 1.3
I don't know why the 30 GB HDD takes 86 seconds for reading a 1 GB
file on 2.5.40 whereas it only took 60 seconds on 2.4.19-pre5. The 120
GB HDD has almost the same reading speed on 2.5.40 (31 seconds) as on
2.4.19-pre5 (32 seconds).
I ran the benchmarks for 2.5.40 just after booting into runlevel 1
(single user mode) on a Suse 8.0 system. The system is a Athlon 700
MHz, KT133 chipset, 256 MB RAM, 256 MB swap.
The "dd" benchmark is:
#!/bin/bash
dd if=/dev/zero of=$1/tmp bs=1000000 &
# a sleep is sometimes necessary for bad performance (to fill cache?)
sleep 30
time dd if=$2 of=/dev/null bs=1000000
Filesystems are:
jfranosc@nomad:~ > mount
/dev/hda3 on / type reiserfs (rw,noatime)
proc on /proc type proc (rw)
devpts on /dev/pts type devpts (rw,mode=0620,gid=5)
/dev/hda1 on /boot type ext2 (rw,noatime)
/dev/hda6 on /home type ext2 (rw,noatime)
/dev/hda7 on /lscratch type reiserfs (rw,noatime)
/dev/hdc2 on /lscratch2 type reiserfs (rw,noatime)
shmfs on /dev/shm type shm (rw)
automount(pid341) on /net type autofs (rw,fd=5,pgrp=341,minproto=2,maxproto=4)
automount(pid334) on /misc type autofs (rw,fd=5,pgrp=334,minproto=2,maxproto=4)
/dev/hdd4 on /mzip type vfat (rw,noexec,nosuid,nodev,user=jfranosc)
/dev/sr0 on /dvd type ext2 (rw,noexec,nosuid,nodev,user=jfranosc)
Boot parameters in /etc/lilo.conf:
append = "hdb=ide-scsi hdd=ide-floppy"
So, kernel 2.5.40 has improved much over 2.4.18 and 2.4.19-pre5, but
there is still room for improvements in this benchmark for up to a
factor of more than 2.
If you have patches that you think should be tested, I'd like to try
them.
Thank you for your attention,
Moritz
--
Dipl.-Phys. Moritz Franosch
http://Franosch.org
On Sun, Oct 06, 2002 at 06:08:00PM +0200, Moritz Franosch wrote:
> Theoretically, the two devices (e.g. the 120 GB HDD and the DVD-RAM in
> nr. 2) are independent (at least when they are on a different IDE
> controller), so the read should perform as good as without the
> additional writing activity (31 seconds, as given in the column
> "expected"). But in reality the read takes 1.8 times as long. With
> 2.4.50, there could be performance gains up to a factor of 2.4 (test
> nr. 4). But that's _much_ better than 2.4.19-pre5 where performance
> gains of up to a factor of 14 (test nr. 2) were possible.
one of the reasons 2.5 works better is that it has more than one bdflush
thread (aka pdflush), so it can submit in parallel to all your storages
and you don't risk to hang on writes on the very slow device, despite
your main harddisk could be completely idle. This isn't easily fixable
in 2.4, it's one of the things 2.5 is there for ;)
Andrea