[Tig] Minimum RAID bandwidth for 2K playback?

Jason Howard jth at spectsoft.com
Fri Jul 18 20:30:16 PDT 2008


Hi Bob-

> A rather bright engineer in another forum has posulated that the path
> to successful real-time 2K playback is to eliminate disk seeks
> entirely. However, I proposed that a few disk seeks are ok given that
> the data can just be read a little faster to make up the difference.

Given that your array is slightly faster than the needed bandwidth, it ends up 
depending mostly on your buffer window.  So for a buffer of 24 frames at 24 
FPS playback, you can't block any longer than 1000ms during any 24 frame 
window.  Put simply, it isn't exactly about removing the seeks alltogether, 
it is about managing them or making sure they don't happen too often.

> I am interesting in hearing what sort of storage hardware is in
> current use for real-time 2K playback and if the ultimate solution was
> to purchase a packaged solution from a proprietary/specialty vendor,
> or to apply sensible engineering and purchase competitive "IT"
> hardware sufficient (capacity and data rate) for the task.
>
> What raw storage bandwidth capacity has been proven to be necessary to
> satisfy the minimum 292MB/second data rate necessary for real-time 2K
> playback?  What sort of RAID topology is used (RAID-0, RAID-5, RAID-6,
> RAID 1+0)?  How many disk drives and what technology (e.g. Ultra-320
> SCSI, SATA, SAS, FC) are they?

We use 3ware (9550/9650/9690) controllers with either SAS or SATA in Raid 0, 
Raid 5, or Raid 1+0.  Arrays are usually 12 drives, but that is usually due 
to capacity.  We have found that the 3ware controllers scale pretty linearly 
(in both reads and writes) until about 10 or 11 drives.

Other than that, there is a bit of tuning in terms of read ahead.  It also 
helps when your application is able to tell the OS when not to cache a 
particular read or write.  As it turns out, a significant bottleneck can be 
the OS in trying to cache (and dump cache) while reading and writing- 
especially since in a typical playback/record scenerio, you won't really need 
that frame again.

Jason

-- 
 Jason Howard

 Professional:  
   SpectSoft
   375 Johnson Ave
   Oakdale, CA 95361, USA
   http://www.spectsoft.com
   jth at spectsoft dot com
   Phone: +1.209.847.7812 Ext.102
   Fax: +1.209.847.7859

 Personal:
   http://www.psinux.org
   jason at psinux.org
   Cell: +1.209.918.3003
   Text Message: jasonsphone at psinux dot org
   IM: jhoward at jabber dot org
   IRC: jhoward at irc dot freenode dot net
   AIM: jh0w4rd



More information about the Tig mailing list