[Asterisk-Users] TDM400P and SCSI/SATA = * noise problems???

[tmassey at obscorp.com]

asterisk-users-bounces at lists.digium.com wrote on 04/19/2005 01:32:57 AM:

> ** Extract begins **
> 
> SCSI RAID can cause the problem.  If disabling hyper threading does not 
> resolve your problem my next suggest would be to revert to a PATA IDE 
> hard drive solution configured to UDMA level 2 using hdparm.  SCSI or 
> SATA causes problems on some systems from what I have seen.  The problem 

> increases when using a SCSI or SATA RAID.
> 
> ** Extract ends **
> 
> I really hope that they are wrong, as I don't feel like throwing away my 

> nice expensive Ultra320 SCSI RAID controller and hot plug drives and 
> replacing them with some crusty old IDE config.  Needless to say I'm not 

> going to go and shell out on IDE controller & drives until I'm a little 
> more certain that this is actually a problem and have asked them for 
> more information.
> 
> Does anyone else find it odd that the TDM could possibly have a problem 
> sharing a box (but not an IRQ) with a SCSI controller?

Yes.  It has to do with latency and bus contention.

I've run a TDM board in an IBM Netfinity 5600 server with an IBM ServeRAID 
3L controller (SCSI-U2W).  The big difference, though, is that the RAID 
controller was on its own PCI bus, and the TDM card was on its own PCI 
bus.

With both controllers on the bus, you can have latency issues.  For 
example, if the RAID controller sets up a DMA of a big chunk of disk, it 
owns the bus for that transfer.  If an Ethernet packet is delayed by 50us 
during that time, nobody cares.  But if the TDM card is delayed, it most 
certainly cares:  especially as its generating 1000 interrupts a second!

That's the problem with the TDM cards.  They do *nothing* on the CPU side. 
 The CPU has to do *everything*, and it has to do it *immediately*.  When 
you are using plain-jane IDE, you can tweak the kernel to put the IDE 
stuff at a low priority.  But when you've got a fancy RAID controller, it 
tends to think it's the most important thing in the system.  And as a 
rule, hard drive I/O usually *is* the most important I/O going on in a 
system.  However, in this case, the TDM card trumps that.  And Digium 
doesn't know how to tweak every last RAID driver in existence for 
low-priority operation--or even if it's possible.  Hence, the 
recommendation for IDE.

> Combined with the fact that they have also recommended that we turn off 
> hyper threading (also causes problems with TDM, apparently), I'm 
> wondering if these cards shouldn't come with a warning not to use 
> anything with half decent performance in your * server!

Yet they require PCI 2.2, which eliminates most Pentium III's and lower! 
:)

I'm still in the midst of testing the TDM cards.  So far, so good, in an 
EPIA-based solution and in the 5600.  But I've been through at least half 
a dozen different systems before I've found these...

Tim Massey