twa driver and 3ware 9690SA issues

[Josh Paetzel]

On Mar 2, 2009, at 4:33 PM, Andrew Snow wrote:

> Josh Paetzel wrote:
>> Which wasn't that much of an issue, as SATA drives aren't capable  
>> of sustained sequential 100 MB/sec transfers anyways
>
> As Scott said, this number is not reflective of actual negotiated  
> rate, and is nothing to worry about.

Great, thanks.

>
>> u0    RAID-10   OK             -       -       64K     596.025    
>> ON     OFF
>> 838860800 bytes transferred in 6.923690 secs (121158052 bytes/sec)
>> Kind of what I expect, and faster than 100 Megs/sec
>
> Well, that's a RAID10 and capable of doing much faster than 120MB/s.
>

Hrmm, I was under the impression that a single read or write would  
only see the sequential transfer rate of one side of the stripe set,  
and that would be fairly close to 120 Megs/sec.

> Sorry, but you cannot expect anywhere near close to the theoretical  
> unformatted maximum media transfer rate of the disks you're testing,  
> especially under FreeBSD and UFS.
>

No need to apologize.

> Testing the raw disks doesn't necessarily help either because then  
> there's no operating system read-ahead or write-caching, which is  
> necessary to lower the latency between access requests.
>
> I don't believe you're getting any SCSI or 3ware problems - just  
> FreeBSD kernel and filesystem issues.
>
> UFS is performance is a bit bizarre: for example I see faster  
> sequential write speeds than read speeds! ZFS improves the situation  
> alot and I get much closer to the theoritical max with my drives.
>
> To give you an idea of how much more efficient ZFS is:  I had a 16  
> disk 3ware SATA RAID6, with UFS, and battery-backed cache.  I  
> converted to a 14 disk ZFS software RAIDZ2 with a smaller 3ware  
> mirrored boot disk.
>
> Quick benchmarked using dd as follows:
>
> Creating a 6GB file:  UFS+3ware=308mb/s ZFS=357mb/s	
> Reading a 6GB file:   UFS+3ware=113mb/s ZFS=424mb/s
>
>
>
> - Andrew


ZFS probably isn't an option for me.  A lot of these systems max out  
at 8 gigs of RAM, and postgresql eats a huge chunk of that, also the  
controllers on the motherboard don't always report back drive failures  
to the OS, as well as have occasional failures detecting new drives  
when we need to hot swap, as well as having my biggest concern, no  
battery backed cache. :(

My main question has been answered, the link speed that the driver  
reports is imaginary.


Thanks,

Josh Paetzel