Hi Alan, hi all
so far, i did some more tests.
First of all, i'm using ext3 filesystem (and it might happen it's very
slow, according with my tests results)
What i did:
Trying ubuntu Hardy LiveCD (to check it's not because
1) im using a hand made compiled kernel
2) i might have removed some package which is actually needed for disk
performances
3) the upgrade screwed something )
So far, from LiveCD i did the same i did before, that is:
root@ubuntu:/media/disk/movies# time cat Lo\ squalo.avi >/dev/null
real 5m16.352s
user 0m0.132s
sys 0m6.244s
root@ubuntu:/media/disk/movies# ls -la Lo\ squalo.avi
-rw-r----- 1 1000 1001 1559521280 2008-03-30 06:37 Lo squalo.avi
it means roughly those old 5mb/sec.
In the meanwhile i checked with top, and all the cpu cycles were taken
by the disk (about 98.0%wa and 0.0%id)
ok so not my fault. Maybe ext3? my volumes looks like:
root@rivendell:~# df
Filesystem 1K-blocks Used Available Use% Mounted on
/dev/sda1 28834716 17088924 10281068 63% /
[...]
/dev/sda3 931618688 613099244 314519444 67% /local
So i got a 30gb partition, and a 900gb partition (roughly).
I tried to copy the file into the smaller partition ( into / )
and:
root@rivendell:/# time cp -rf /local/movies/Lo\ squalo.avi /
real 6m21.978s
user 0m0.150s
sys 0m18.170s
root@rivendell:/# time cat /Lo\ squalo.avi >/dev/null
real 1m33.913s
user 0m0.150s
sys 0m5.370s
root@rivendell:/# time cat /local/movies/Lo\ squalo.avi >/dev/null
real 6m13.549s
user 0m0.230s
sys 0m5.930s
During this tests, the first and the last command still showed like 98.0%wa
while with the second, (reading the file from the / fs) i had about
50-60%wa and some idle cycles (about 20% or so, sometimes 0% but not
always) with about 16mb/sec transfer rate (which is still slow, to
me.. again, hdparm -tT shows:
root@rivendell:/# hdparm -tT /dev/sda
/dev/sda:
Timing cached reads: 476 MB in 2.00 seconds = 237.68 MB/sec
Timing buffered disk reads: 220 MB in 3.02 seconds = 72.80 MB/sec )
This test shows to me that it might be a filesystem issue, but still,
is it possible it's so expensive? i mean, even from a 30gb partition
(which is the norm) reading operations take almost all the cpu power
(even if it's finally faster).
Ok, my cpu is not lighting fast (AthlonXP 800mhz) but i had almost
better performances on a pentiumII 350mhz and 320gb IDE hard drive
(debian testing on ext3).
do you have any clue about what could i do to improve fs performances?
Thank you in advance and excuse me for the boring mail!
Paolo
On 5/1/08, Alan Cox [off-list ref] wrote:
On Thu, 1 May 2008 21:16:24 +0200
Paolo [off-list ref] wrote:
quoted
Umh..
Thank you for your answer!
Actually, i think the same. Hdparm tests shows good results.
hdparm results mean the disk setup is ok
quoted
what i dont get is why the cpu load raise to keep idle to 0% like if
it was pio mode, even if it is in udma6..
What else could be, other than the libata layer not setted in dma
(which is not)?
Anything above the disk driver layer - file system, block queue setup
or something happening which causes a lot of I/O in small chunks (which
ruins a disks performance). This is usually distro level stuff so best
analysed by the distribution people.
quoted
ps: no, im not using LVM. But ubuntu is using their uuid stuff in
fstab to boot which i quite dont know at all..
maybe that?
No that wouldn't explain it at all. The uuid/fstab stuff is just boot
time stuff so doesn't get in the way.
What file system are you using ?
Actually, i was curious to see how my old Pentium2 performs, so i
tried an ubuntu gutsy on it.
Surprising results:
root@erebor:~# df
Filesystem 1K-blocks Used Available Use% Mounted on
/dev/sda1 14421344 2177664 11511120 16% /
[...]
/dev/sda3 273522480 260268272 12230208 96% /local
(so one 15gb partition, and one 270gb partition, both ext3)
hdparm -tT /dev/sda
/dev/sda:
Timing cached reads: 158 MB in 2.00 seconds = 78.93 MB/sec
Timing buffered disk reads: 84 MB in 3.03 seconds = 27.74 MB/sec
crappy results, BUT:
root@erebor:~# time cat /local/SAT/aba/bigfile.avi >/dev/null
real 0m24.637s
user 0m0.236s
sys 0m5.676s
root@erebor:~# ls -la /local/SAT/aba/bigfile.avi
-rw-rw-rw- 1 root root 732903424 2007-05-29 22:43 /local/SAT/aba/bigfile.avi
which means about 28mb/sec (what hdparm -t state). Here i still get
0%id cpu, but at least it goes fast. Now: this machine has an intel
i440bx chipset.
The other one a Via KT600.
Both of them run or ran ubuntu gutsy (now the new one has hardy, but
before , with gutsy, i had same results)
So... is sata_via buggy with via VT8237 soutbridge?
On 5/17/08, Paolo [off-list ref] wrote:
Hi Alan, hi all
so far, i did some more tests.
First of all, i'm using ext3 filesystem (and it might happen it's very
slow, according with my tests results)
What i did:
Trying ubuntu Hardy LiveCD (to check it's not because
1) im using a hand made compiled kernel
2) i might have removed some package which is actually needed for disk
performances
3) the upgrade screwed something )
So far, from LiveCD i did the same i did before, that is:
root@ubuntu:/media/disk/movies# time cat Lo\ squalo.avi >/dev/null
real 5m16.352s
user 0m0.132s
sys 0m6.244s
root@ubuntu:/media/disk/movies# ls -la Lo\ squalo.avi
-rw-r----- 1 1000 1001 1559521280 2008-03-30 06:37 Lo squalo.avi
it means roughly those old 5mb/sec.
In the meanwhile i checked with top, and all the cpu cycles were taken
by the disk (about 98.0%wa and 0.0%id)
ok so not my fault. Maybe ext3? my volumes looks like:
root@rivendell:~# df
Filesystem 1K-blocks Used Available Use% Mounted on
/dev/sda1 28834716 17088924 10281068 63% /
[...]
/dev/sda3 931618688 613099244 314519444 67% /local
So i got a 30gb partition, and a 900gb partition (roughly).
I tried to copy the file into the smaller partition ( into / )
and:
root@rivendell:/# time cp -rf /local/movies/Lo\ squalo.avi /
real 6m21.978s
user 0m0.150s
sys 0m18.170s
root@rivendell:/# time cat /Lo\ squalo.avi >/dev/null
real 1m33.913s
user 0m0.150s
sys 0m5.370s
root@rivendell:/# time cat /local/movies/Lo\ squalo.avi >/dev/null
real 6m13.549s
user 0m0.230s
sys 0m5.930s
During this tests, the first and the last command still showed like 98.0%wa
while with the second, (reading the file from the / fs) i had about
50-60%wa and some idle cycles (about 20% or so, sometimes 0% but not
always) with about 16mb/sec transfer rate (which is still slow, to
me.. again, hdparm -tT shows:
root@rivendell:/# hdparm -tT /dev/sda
/dev/sda:
Timing cached reads: 476 MB in 2.00 seconds = 237.68 MB/sec
Timing buffered disk reads: 220 MB in 3.02 seconds = 72.80 MB/sec )
This test shows to me that it might be a filesystem issue, but still,
is it possible it's so expensive? i mean, even from a 30gb partition
(which is the norm) reading operations take almost all the cpu power
(even if it's finally faster).
Ok, my cpu is not lighting fast (AthlonXP 800mhz) but i had almost
better performances on a pentiumII 350mhz and 320gb IDE hard drive
(debian testing on ext3).
do you have any clue about what could i do to improve fs performances?
Thank you in advance and excuse me for the boring mail!
Paolo
On 5/1/08, Alan Cox [off-list ref] wrote:
quoted
On Thu, 1 May 2008 21:16:24 +0200
Paolo [off-list ref] wrote:
quoted
Umh..
Thank you for your answer!
Actually, i think the same. Hdparm tests shows good results.
hdparm results mean the disk setup is ok
quoted
what i dont get is why the cpu load raise to keep idle to 0% like if
it was pio mode, even if it is in udma6..
What else could be, other than the libata layer not setted in dma
(which is not)?
Anything above the disk driver layer - file system, block queue setup
or something happening which causes a lot of I/O in small chunks (which
ruins a disks performance). This is usually distro level stuff so best
analysed by the distribution people.
quoted
ps: no, im not using LVM. But ubuntu is using their uuid stuff in
fstab to boot which i quite dont know at all..
maybe that?
No that wouldn't explain it at all. The uuid/fstab stuff is just boot
time stuff so doesn't get in the way.
What file system are you using ?
root@ubuntu:/media/disk/movies# ls -la Lo\ squalo.avi
-rw-r----- 1 1000 1001 1559521280 2008-03-30 06:37 Lo squalo.avi
it means roughly those old 5mb/sec.
In the meanwhile i checked with top, and all the cpu cycles were taken
by the disk (about 98.0%wa and 0.0%id)
98.0%wa - wait time, not CPU time.
Its still a very slow copy rate unless the disks are on the same IDE
channel. How is the disk set up - ext3 direct or via raid/lvm ?
Also is a copy of the movie to /dev/null slow or fast, ditto a copy
from /dev/zero to disk of that sort of size ?
Alan
So, let me clarify a bit:
i did the same test in 3 different configuration:
"cat bigfile >/dev/null"
3 configuration were
1) Machine A (the slow one) from the big partition (1000gb) - got 5mb/s
2) Machine A from the small partition - got 16mb(s
3) Machine B from 300gb partition - got 28mb/s
Machine A: AthlonXP 800 -Via KT600 - SATA disk western digital -
direct ext3 filesystems (no raid-no lvm-no nothing)
Machine B: Pentium 2 350 - i440bx - seagate IDE drive - direct ext3 filesystems
I also did another test: putting a cpu intensive process in the
background with nice -n 20. The test showed the process was not being
executed almost at all, so those %wa cpu cycles are somehow cpu cycles
since they dont let other processes execute. (in fact when i do stuff
with the disk my system became incredibly slow)
Finally i tried the test you suggested me
root@rivendell:~# time dd if=/dev/zero of=/tmp/testfile bs=1024000 count=1000
1000+0 records in
1000+0 records out
1024000000 bytes (1,0 GB) copied, 16,9609 s, 60,4 MB/s
real 0m17.119s
user 0m0.010s
sys 0m12.890s
with a sample cpu usage of:
Cpu(s): 0.7%us, 15.0%sy, 0.0%ni, 76.1%id, 6.6%wa, 0.0%hi, 1.7%si, 0.0%st
root@rivendell:~# time dd if=/dev/zero of=/local/testfile bs=1024000 count=1000
1000+0 records in
1000+0 records out
1024000000 bytes (1,0 GB) copied, 23,3957 s, 43,8 MB/s
real 0m23.898s
user 0m0.010s
sys 0m14.090s
with a sample cpu usage of
Cpu(s): 1.3%us, 50.7%sy, 0.0%ni, 0.0%id, 45.4%wa, 0.7%hi, 2.0%si, 0.0%st
( / was the 30gb partition, /local the 900gb)
On 5/18/08, Alan Cox [off-list ref] wrote:
quoted
real 5m16.352s
user 0m0.132s
sys 0m6.244s
So the CPU usage is ok
quoted
root@ubuntu:/media/disk/movies# ls -la Lo\ squalo.avi
-rw-r----- 1 1000 1001 1559521280 2008-03-30 06:37 Lo squalo.avi
it means roughly those old 5mb/sec.
In the meanwhile i checked with top, and all the cpu cycles were taken
by the disk (about 98.0%wa and 0.0%id)
98.0%wa - wait time, not CPU time.
Its still a very slow copy rate unless the disks are on the same IDE
channel. How is the disk set up - ext3 direct or via raid/lvm ?
Also is a copy of the movie to /dev/null slow or fast, ditto a copy
from /dev/zero to disk of that sort of size ?
Alan