Problem with concurrent SATA-Writes

Problem with concurrent SATA-Writes

[Tobias Reinhard]

Hi All!

I have a problem with two SATA-Discs. I have an onboard SIL3114 with 
four SATA-Ports an onboard NVIDIA with two SATA-Ports - both controllers 
are disable via BIOS (and are not detected by Linux)(only for this test 
of course - normally I have other HDD on this ports). The only 
controller that is found is the add-on controller in a PCI-Slot 
(SIL3112). And that is the one I have trouble with.

If I read or write  (via dd) from the first one -> no problems. Same 
when read or write from the second or when I read (only read!) from both 
at the same time. Data-Transfer-Rate is around 45MB/s for one HDD.

The problem occures when I try to write on both discs at the same time. 
For example I write /dev/zero to the first one and then start to write 
/dev/zero to the second one. The System-Load goes up to 4 with nearly 
100% io-wait and nearly no really write-access to the drives.

Any hints?

- no errormessages in syslog
- happens with Kernel 2.6.11.7 and with 2.6.12-rc5
- HDDs are Samsung Spinpoint 200GB
- (I use the SCSI-SATA-Drivers)
- anything else you need?

Thanks

Tobias

Problem with concurrent SATA-Writes

[Tobias Reinhard]

I have a problem with two SATA-Discs. I have an onboard SIL3114 with
four SATA-Ports an onboard NVIDIA with two SATA-Ports - both controllers
are disable via BIOS (and are not detected by Linux)(only for this test
of course - normally I have other HDD on this ports). The only
controller that is found is the add-on controller in a PCI-Slot
(SIL3112). And that is the one I have trouble with.

If I read or write  (via dd) from the first one -> no problems. Same
when read or write from the second or when I read (only read!) from both
at the same time. Data-Transfer-Rate is around 45MB/s for one HDD.

The problem occures when I try to write on both discs at the same time.
For example I write /dev/zero to the first one and then start to write
/dev/zero to the second one. The System-Load goes up to 4 with nearly
100% io-wait and nearly no write-access to the drives.

Any hints?

- no errormessages in syslog
- happens with Kernel 2.6.11.7 and with 2.6.12-rc5
- HDDs are Samsung Spinpoint 200GB
- (I use the SCSI-SATA-Drivers)

Tobias

Re: Problem with concurrent SATA-Writes

[Lennart Sorensen]

On Tue, May 31, 2005 at 07:16:42PM +0200, Tobias Reinhard wrote:
> I have a problem with two SATA-Discs. I have an onboard SIL3114 with
> four SATA-Ports an onboard NVIDIA with two SATA-Ports - both controllers
> are disable via BIOS (and are not detected by Linux)(only for this test
> of course - normally I have other HDD on this ports). The only
> controller that is found is the add-on controller in a PCI-Slot
> (SIL3112). And that is the one I have trouble with.
> 
> If I read or write  (via dd) from the first one -> no problems. Same
> when read or write from the second or when I read (only read!) from both
> at the same time. Data-Transfer-Rate is around 45MB/s for one HDD.
> 
> The problem occures when I try to write on both discs at the same time.
> For example I write /dev/zero to the first one and then start to write
> /dev/zero to the second one. The System-Load goes up to 4 with nearly
> 100% io-wait and nearly no write-access to the drives.
> 
> Any hints?
> 
> - no errormessages in syslog
> - happens with Kernel 2.6.11.7 and with 2.6.12-rc5
> - HDDs are Samsung Spinpoint 200GB
> - (I use the SCSI-SATA-Drivers)

Might it be a problem with having two things using /dev/zero at the same
time?

What blocksize do you use with dd?

What happens if you do dd from /dev/zero to two different files on one
of the hds (with some filesystem on the drive obviously)?

How about:
dd if=/dev/zero bs=128k| tee >(dd of=/dev/disk1 bs=128k) | dd of=/dev/disk2 bs=128k

That seems pretty fast here (writing to two files in /tmp rather than
two disks since I don't have two disks I can just destroy.).

Len Sorensen

Re: Problem with concurrent SATA-Writes

[Tobias Reinhard]

Lennart Sorensen wrote:
 > Might it be a problem with having two things using /dev/zero at the same
 > time?
No, dding from /dev/zero two times at the same time is working perfectly

 > What blocksize do you use with dd?
Generally I use 1M. But the Problem is the same with 128k or with 4k

 > What happens if you do dd from /dev/zero to two different files on one
 > of the hds (with some filesystem on the drive obviously)?
I did it without a FS on it and its no problem to write at two different 
locations - as long as they are on one disc.

It seems to me that the driver has problems with handling two 
write-requests on two ports at the same time.

Tobias

Problem with concurrent SATA-Writes

[Tobias Reinhard]

Hello Jeff, hello to the linux-ide-list!

I already posted this to the lkml, but unfortunately I got no usefull 
answer. I think the problem is related to the SIL-Driver so maybe you 
are the right person to contact - or someone at the linux-ide-list...

If you need any other informations or logs please tell me. I really want 
to use these discs.

Tobias

Original Msg:
--------
I have a problem with two SATA-Discs. I have an onboard SIL3114 with
four SATA-Ports an onboard NVIDIA with two SATA-Ports - both controllers
are disable via BIOS (and are not detected by Linux)(only for this test
of course - normally I have other HDD on this ports). The only
controller that is found is the add-on controller in a PCI-Slot
(SIL3112). And that is the one I have trouble with.

If I read or write  (via dd) from the first one -> no problems. Same
when read or write from the second or when I read (only read!) from both
at the same time. Data-Transfer-Rate is around 45MB/s for each HDD.

The problem occures when I try to write on both discs at the same time.
For example I write /dev/zero to the first one and then start to write
/dev/zero to the second one. The System-Load goes up to 4 with nearly
100% io-wait and nearly no write-access to the drives.

Any hints?

- no errormessages in syslog
- happens with Kernel 2.6.11.7 and with 2.6.12-rc5
- HDDs are Samsung Spinpoint 200GB
- (I use the SCSI-SATA-Drivers)

Re: Problem with concurrent SATA-Writes

[Jeff Garzik]

Tobias Reinhard wrote:
> Hello Jeff, hello to the linux-ide-list!
> 
> I already posted this to the lkml, but unfortunately I got no usefull 
> answer. I think the problem is related to the SIL-Driver so maybe you 
> are the right person to contact - or someone at the linux-ide-list...
> 
> If you need any other informations or logs please tell me. I really want 
> to use these discs.

Yes, I think this is the third time I've read your email.  You can stop 
posting it now.

Sometimes we just don't have time to read and respond to every support 
request.  If you absolutely require support, please use a paid support 
channel like SuSE or Red Hat.

	Jeff

Re: Problem with concurrent SATA-Writes

[Tobias Reinhard]

Jeff Garzik wrote:
> Yes, I think this is the third time I've read your email.  You can stop 
> posting it now.
> 
> Sometimes we just don't have time to read and respond to every support 
> request.  If you absolutely require support, please use a paid support 
> channel like SuSE or Red Hat.
Sorry if I annoyed or bothered You or any other one on the list.

I didn't expect that You or any other immediately care about my problems 
and solve them or already have a solution. I could try to fix the driver 
myself but as it seems to be a SIL-specific Problem I need the help or 
some hints from someone who has the (nda-)docs. Maybe they already noted 
this as an errata for which some kind of quirk is needed.

I worked-around the problem by putting the discs on a different 
controller. It seems that this only happens with these Samsung-drives on 
this controller.

Just in case you face this or a similar Problem in the future or find 
the time to investigate it...

BTW: with the siimage driver the transfer-rate is low (20MB/s instead of 
45MB/s) but at least it's working.

BTW2: same with 2.6.12-rc5-mm2


	Tobias

Re: Problem with concurrent SATA-Writes

[Jeff Garzik]

Tobias Reinhard wrote:
> 
> 
> Jeff Garzik wrote:
> 
>> Yes, I think this is the third time I've read your email.  You can 
>> stop posting it now.
>>
>> Sometimes we just don't have time to read and respond to every support 
>> request.  If you absolutely require support, please use a paid support 
>> channel like SuSE or Red Hat.
> 
> Sorry if I annoyed or bothered You or any other one on the list.
> 
> I didn't expect that You or any other immediately care about my problems 
> and solve them or already have a solution. I could try to fix the driver 
> myself but as it seems to be a SIL-specific Problem I need the help or 
> some hints from someone who has the (nda-)docs. Maybe they already noted 
> this as an errata for which some kind of quirk is needed.

The sata_sil hardware docs are public:
http://gkernel.sourceforge.net/specs/sii/3112A_SiI-DS-0095-B2.pdf.bz2

	Jeff