FreeBSD 5.3b7and poor ata performance

[Scott Long]

Charles Swiger wrote:
> On Oct 25, 2004, at 2:29 PM, Scott Long wrote:
> 
>>> Also, there is an unresolvable question. Why two 52MB/s disks
>>> in raid0 has a throughput of 40MB/s and for raid1 18MB/s??
>>
>>
>> Would you _PLEASE_ stop trying to associate RAID with performance!
>> RAID is about reliability and reduncdancy, not about speed.
> 
> 
> All RAID modes make tradeoffs between performance, reliability, and cost.
> 
> RAID-1 mirroring and RAID-5 provide higher reliability by using partial 
> or full redundancy.  However, RAID-0 striping provides no additional 
> reliability: the primary reason for using RAID-0 is to improve 
> performance by accessing two or more devices in parallel.
> 
>> Some cases can give you desirable performance increases as a side effect,
>> but that is not the primary goal.
> 
> 
> Disagree.  Why else would you use RAID-0 striping?
> 
> [ If you simply want to create a logical volume bigger than the size of 
> a physical drive, you can use concatenation instead. ]
> 
>> Specifically in this case, the
>> GEOM raid classes are fairly new and have not had the benefit of
>> years of testing.  I'd much rather that the focus be on stability
>> and reliability for them, not speed.  Once the primary goals of
>> RAID are satisfied then we can start looking at performance.
> 
> 
> Your position is certainly reasonable: if a storage system is not 
> reliable, how fast it performs is something of a moot point.  :-)  
> However, this being said, a RAID-0 implementation needs to improve 
> performance compared with using a bare drive if it is to be useful.
> 

Well, RAID-0 is a special case =-)  That said, putting discrete RAID
classes into the GEOM layer is something of a new adventure, so I'm
not surprised to hear about performance problems, even in RAID-0.
There might be extra data copies or path latencies that weren't planned
for or expected.  It's definitely something to look at.  But it's also
a very new subsystem, so it would be unfair to judge FreeBSD performance
with it.

Scott