[Gllug] A SATA Raid card worth having?

Nix nix at esperi.org.uk
Fri Feb 3 21:26:54 UTC 2006


On Thu, 2 Feb 2006, Christian Smith stated:
> On Thu, 2 Feb 2006, Martin A. Brooks wrote:
> 
>>Nix wrote:
>>
>>>Of course if you're doing small piecemeal writes then there'll be a lot
>>>of extra reads to compute parity across all stripes... but small
>>>piecemeal writes are a fairly uncommon thing.
> 
> No extra reads necassary. XOR is commutative, so you just XOR the old
> block against the parity block, then XOR the new block against the parity
> block and write them both.

Ooh. True. Didn't think of that. :)

>>Apart from mail servers, database servers, web caching servers, log
>>servers etc etc, yes, they're fairly uncommon :)

Mail servers: not piecemeal, generally either appends (mbox) or new
files (maildir, nnmh, nnml...).

Web caching servers: generally create whole files as a lump.

Log servers: plain appends.

Database servers: the elephant in the room, these guys really *do*
scatter writes around like anything.

> Against the cost of actually writing the data to disk, is the time to do
> an XOR across a few 1K blocks ever going to be significant? Small writers
> like the above applications are almost always seek bound. A seek bound
> disk will handle perhaps 200-300 transactions per second? At 1K per
> transaction, that's less than 1/2 MBps per disk.

Quite so. And the big writes will end up saturating the PCI bus...

-- 
`I won't make a secret of the fact that your statement/question
 sent a wave of shock and horror through us.' --- David Anderson
-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug




More information about the GLLUG mailing list