2120S poor performance

[Jin Guojun [VFFS]]

Mike Horwath wrote:

>On Wed, Dec 08, 2004 at 11:51:21AM -0500, Bill Anderson wrote:
>  
>
>>I've got a 3x72g RAID5 array with U320 disks on an Adaptec 2120S
>>controller under Freebsd 4.10.  I'm getting about 25MB/s for sequential
>>reads/writes (e.g.  dd if=/dev/aacd0s1b of=/dev/null bs=32k).  I've
>>turned on write caching for the container (container set cache
>>/write_cache_enable) and turned off the read caching, per previous posts
>>to this list.  I'm going to try setting hw.aac.iosize_max to 96k tonight
>>to see if that helps.  I'm currently using an IDE disk that gets about
>>45MB/s, so I can hardly consider it an "upgrade" to switch to a 25MB/s
>>U320 SCSI system.  :/
>>
>>I've also read that the 2120S is slow because of its design.  I'm trying
>>to figure out whether the performance can be increased significantly (I
>>saw a posting of a linux user getting 37.5MB/s, which although still slow
>>might be acceptable), or if I'm better off getting a new card.  I'm
>>thinking the latter is the case based on previous postings, but since it
>>will probably be quite expensive to replace it, I wanted to get some more
>>data.
>>
>>What performance should I be expecting from a decent U320 RAID5
>>controller?
>>
>>Has anyone gotten a 2120S to perform above 30Mb/s in FreeBSD?
>>
>>What's the cheapest controller that still gives reasonable performance?
>>(If you could give a couple different ones, with their associated
>>performance (under FreeBSD), or tell me where to find such information,
>>that would be great)
>>    
>>
>
>I know I left quite a bit of cruft above, but I don't see these kinds
>of performance numbers you are seeing.
>
>First, SCSI does a lot with overlapping I/O and tagged queuing, only
>the newest IDE (SATA) systems are starting to support such things.
>
>Second, SCSI disks run at a faster RPM most of the time (only the WD
>Raptor hits the 10K mark), lowering seek time significantly.
>
>Third, my numbers:
>
>31MB/sec using
>
>	dd if=/dev/aacd0s1b of=/dev/null bs=32k count=20000
>
>34MB/sec using
>
>	dd if=/dev/aacd0s1b of=/dev/null bs=64k count=20000
>
>33MB/sec using
>
>	dd if=/dev/aacd0s1b of=/dev/null bs=128k count=20000
>
I do not remember where the original problem of 2120s was posted.
The read in RAID 5 mode is 40-56 MB/s, where write rate is 5-13 MB/s.
The most problem is at writing to RAID, average around 8 MB/s.
SCSI buses self is Okay because doing multiple R/W without RAID
can get maximum disk I/O. Soft RAID, like Vinum, also slows down
disk array performance. SO, it sounds like that RAID design has
big overhead to kill the performance.

dd if=/dev/aacd0s1b of=/dev/null bs=128k count=20000
dev/null bs=128k count=20000
20000+0 records in
20000+0 records out
2621440000 bytes transferred in 61.651078 secs (42520587 bytes/sec)

Adaptec guy told me that this is a design issue. One older model,
which may be out of market, is performing better, but I would not
expect a large difference as soft RAID do not give better performance.

-- 
------------ Jin Guojun ----------- v --- jin at george.lbl.gov ---
Distributed Systems Department		http://www.dsd.lbl.gov/~jin
Lawrence Berkeley National Laboratory,	Berkeley, CA 94720