[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: sas fileserver



on Sat, Jan 11, 2003 at 03:34:55PM -0800, Michael West (quagly@mitzit.net) wrote:
> On Fri, Jan 10, 2003 at 10:45:51AM +0000, Karsten M. Self wrote:
> > on Tue, Jan 07, 2003 at 06:21:17PM -0800, Michael West (web@mitzit.net) wrote:
> > > I have been asked to help with getting a server for SAS.  One of
> > > the large expenses of this is the 200Gb+ RAID-5 disk on the EMC
> > > frame.  When presented with $$$$$ the question came, can't I just
> > > get something I can put under my desk and save $$$$$? 

<...>

> > Michael, a few suggestions.
> > 
> > I've done a lot of SAS work, most of it in my past.  I've also
> > worked with GNU/Linux and some RAIDed filestorage, as well as Samba,
> > more recently.  GNU/Linux and Samba should be more than robust
> > enough for this purpose.
> > 
> > First, if what you're replacing is an EMC server, I'd suggest going
> > whole-hog with GNU/Linux:  SCSI RAID beats software on performance,
> > and IDE RAID on reliability.  The cost is higher by a significant
> > fraction (more than double), but if this is your primary data store,
> > that shouldn't be a hard sell.  200 GiB isn't all that big these
> > days (you can buy single IDE drives with that capacity).  Focus on
> > reliability and backups.  I've had very mixed results with 3Ware's
> > Escalade products (5xxx, 6xxx, and 7xxx) over a couple of years.

> The cost of a SCSI RAID adapter and SCSI vs. IDE disk is tiny compared
> to what we get charged to use the EMC.  
> 
> Is there an adapter you would recommend?

Hmm...  Some years back, LJ ran a shootout which showed Adaptec was
highly overpriced based on performance.  I don't have recent SCSI
details, but would suggest speccing a high performance RAID, and asking
around from the vendors near the top of that pile.

> > SAS analysis usage is usually a large single data pull, followed by
> > summarization and/or subsetting.  Networked access kills
> > performance, so you're likely not going to have all that much
> > traffic on the dataserver.  If you can run multiple NICs out of the
> > box, either dedicated to a single analyst's PC, or on a
> > load-balanced network, you'll improve throughput markedly.
> > Contention on the fileserver itself is likely to be low, but SCSI
> > will help you there.
>
>      Dual attached network.  Got it.
> 
> > The pessimal configuration is when your SAS programmers try to do
> > *all* their work on the fileserver, and there's always some yahoo
> > who does.  Saving working sets back is reasonable, but using the
> > server for SASWORK, SASSSORT, or other temporary or scratch space,
> > really loads up network traffic.  Discourage this if possible.
> 
>      We will have a large SASWORK on the application server.  A
>      SAS/Compaq rep is going to come out and train our SAS folks on how
>      to get the processing going on in the right place.  

My SAS config experience is more with Unix than WinXX, but essentially
you're looking at the SASWORK and SASSORT definitions.  If you can
enforce these via policy  (and I'm not sure you can), that's your safest
bet. Otherwise, just ID the hot cat5 cables by touch ;-)

Peace.

-- 
Karsten M. Self <kmself@ix.netcom.com>        http://kmself.home.netcom.com/
 What Part of "Gestalt" don't you understand?
   GNU/Linux web browsing mini review:  Galeon.  Kicks ass.
     http://galeon.sourceforge.org/



Reply to: