[Gllug] IDE RAID

[Christian Smith]

On Thu, 12 Jun 2003, Chris Bell wrote:

>On Thu 12 Jun, Simon A. Boggis wrote:
>> 
>
>> 
>> IMHO SCSI always seems much faster compared to IDE when there is lots of
>> reading and writing going on all over the place by multiple processes
>> (such as in a shared fileserver).
>> 
>> Simon
>> 
>> 
>   Is that because the SCSI system is able to disconnect from a drive
>between sending a command and returning to continue, allowing the drive seek
>time, and enabling the processor to access other drives during the wait? The
>same should be possible with only one IDE drive per controller.

Typical drive performance comes down to fast it can do random seeks 
(faster seeks means more transactions,) rather than data throughput. 
Transferring the data over the bus is actually not the significant 
component of a transaction, seeking to the correct cylinder and waiting 
for the data to pass under the head is the bottleneck. It takes typically 
8.5ms to do this on a 7200 RPM drive, yet only 0.47ms to transfer 64k of 
data over a 133MB/s ATA bus.

As SCSI drives are typically in the 10k-15k spindle speed range, they have 
lower average seek times, and hence blow away IDE drives on the number of 
transactions per second they can do.

Another performance tip based on that, forget striping on a system where
random access is prevelent. You're better off distributing swap files and
file systems across spindles, so unrelated transactions can occur in
parallel, instead of in series as it would with striping. An example is a
database server, with different table's files on different disks, allowing
parallel access to both tables.

An NFS server would also benefit from SCSI, as NFS writes[0] requires data
to be committed to disk before completing.

Christian

[0] NFSv3 changed that?

-- 
"You have not converted a man because you have silenced him." -John Morley

----------------------------------------------------------------------
    /"\ 
    \ /    ASCII RIBBON CAMPAIGN - AGAINST HTML MAIL 
     X                           - AGAINST MS ATTACHMENTS
    / \ 


-- 
Gllug mailing list  -  Gllug at linux.co.uk
http://list.ftech.net/mailman/listinfo/gllug