Testing new mpr driver

[Kenneth D. Merry]

On Wed, May 07, 2014 at 10:59:14 +0200, Borja Marcos wrote:
> 
> Hi
> 
> I just saw that there's a new driver for LSI3008 SAS3 cards. I have updated a test system to -STABLE, installed one of those cards (in this case, the OEM version sold by IBM as a HBA, with IT firmware).
> 
> mpr0: <LSI SAS3008> port 0x3f00-0x3fff mem 0x912f0000-0x912fffff irq 32 at device 0.0 on pci17
> 
> 
> 
> 
> Excellent so far I have noticed a difference with the mps based LSI2008 cards: in the same hardware configuration as my previous tests (except, of course, the HBA) it seems to be faster. And I am using just one cable to connect the HBA to the backplane instead of two, I didn't have another cable available.
> 
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 25 lun 0 (pass0,da0)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 26 lun 0 (pass1,da1)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 27 lun 0 (pass2,da2)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 28 lun 0 (pass3,da3)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 29 lun 0 (pass4,da4)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 30 lun 0 (pass5,da5)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 33 lun 0 (pass6,da6)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 34 lun 0 (pass7,da7)
> <ATA Samsung SSD 840 BB0Q>         at scbus0 target 35 lun 0 (pass8,da8)
> <IBM-ESXS SAS EXP BP 61A6>         at scbus0 target 263 lun 0 (pass9,ses0)
> <IBM-ESXS SAS EXP BP 61A6>         at scbus0 target 288 lun 0 (pass10,ses1)
> 
> I have created a ZFS pool with the SSDs. With the LSI2008 based card I needed to run several bonnie++ instances in parallel in order to saturate the I/O bandwidth, and with this card a single instance reaches peak performance. 
> 
> Does it make sense at all?

That's hard to say.  If you're using a 6Gb expander, you would have half of
the available SAS bandwidth if you only connected four lanes from the
controller to the expander instead of 8.  If you somehow have a 12Gb
expander (it isn't obvious from the model number above what the expander
speed is), then you would have the same amount of bandwidth.

One thing that could be happening is you may have lower latency through the
new 12Gb controller.

> I assume that the driver could be considered almost production ready. Is it going to be included in 9.3 as well, as the manpage states, or is it a typo?
> 

It is the same driver that LSI has had on their web site, but brought up to
date with all of the changes that have gone into mps(4) lately.

We're planning to merge it to stable/9 so it will go into 9.3, but there
were some bugs in stable/9 support that I fixed in the past couple of days.
We have to wait for the 3 day waiting period to expire before merging the
driver and all the fixes into stable/9.

By the way, if you run with INVARIANTS enabled, you may run into some
issues (i.e. a panic) on reboot until we merge r265485 to stable/10.

Ken
-- 
Kenneth Merry
ken at FreeBSD.ORG