Other Issue in MPR or MPT driver?

[nmath]

Hi,

Currently in our development environment for storage, we have 2 different servers, one equipped with LSI SAS card and other with LSI card for SATA drives.

When we run the following command:
dd if=/dev/zero of=/dev/da10 bs=8192 count=131072

Basically, we are writing raw IO of size 1 GB on /dev/da10. The write operation of 1 GB takes ~20 minutes. However, when we perform the same operation on a Linux CentOS 7, the write operation takes only 11 seconds.

We got the same result on FreeBSD 10.1, 10.2 & current 11.0 version.

After suspecting MPR & MPT driver, we further run the same dd(8) command on a system which has the SATA connected through AHCI interface, in this case it took 24 seconds.

Other interesting behavior observed is:

# newfs /dev/da10
#mount /dev/da10   /mnt/tmp
#dd if=/dev/zero of=/dev/da10 bs=8192 count=131072

In this case it took 7 seconds, considering that FS cache comes into play and enables the write so fast. Immediately after the completion of the write, we unmounted the drive.

#umount /mnt/tmp

Considering that the flush will be kicked in and expected to take around ~20 minutes, but, it took only 7 seconds.

Are there any known performance issues in mpr(4) & mpt(8) drivers?

It would be helpful to understand the issue.

Thanks and Regards,
Nagaraj

 

[nmath]

Hi,

Currently in our development environment for storage, we have 2 different servers, one equipped with LSI SAS card and other with LSI card for SATA drives.

When we run the following command:
dd if=/dev/zero of=/dev/da10 bs=8192 count=131072

Basically, we are writing raw IO of size 1 GB on /dev/da10. The write operation of 1 GB takes ~20 minutes. However, when we perform the same operation on a Linux CentOS 7, the write operation takes only 11 seconds.

We got the same result on FreeBSD 10.1, 10.2 & CURRENT 11.0 version.

After suspecting mpr(4) & mpt(4) driver, we further run the same dd(1) command on a system which has the SATA connected through AHCI interface, in this case it took 24 seconds.

Other interesting behavior observed is:

• # newfs /dev/da10
• # mount /dev/da10 /mnt/tmp
• # dd if=/dev/zero of=/dev/da10 bs=8192 count=131072

In this case it took 7 seconds, considering that FS cache comes into play and enables the write so fast. Immediately after the completion of the write, we unmounted the drive.

• # umount /mnt/tmp

Considering that the flush will be kicked in and expected to take around ~20 minutes, but, it took only 7 seconds.

Are there any known performance issues in mpr(4) & mpt(4) drivers?

It would be helpful to understand the issue.

Thanks and Regards,
Nagaraj

 

[SirDice]

I can't speak for mpr(4) but we have several systems running on mpt(4). Including a rather heavily loaded MySQL server. We don't have any performance issues, everything is, actually, running faster than was initially expected.

Make sure you measure both systems the same way. One thing that came to mind is setting the card's cache to Write-Back instead of Write-Through. The latter would be a lot slower.

 

[nmath]

SirDice,
Thanks for the reply, we are measuring both the systems the same way. I will check the setting regarding the card's cache.
One thing I forgot to mention is that, we are running FreeBSD in dom0 of Xen environment, do yo suspect anything would be related to this?


Thanks and Regards,
Nagaraj

 

[SirDice]

One thing I forgot to mention is that, we are running FreeBSD in dom0 of Xen environment, do yo suspect anything would be related to this?

That's possible. As far as I know FreeBSD dom0 is still experimental and requires a fairly recent 11-CURRENT.