[Asterisk-Users] RAID affecting X100P performance...

[Scott Laird]

On Jul 21, 2004, at 4:25 PM, Kevin P. Fleming wrote:

> Scott Laird wrote:
>
>> That hasn't been my experience at all.  Frankly, I've never seen a 
>> cheap (<$3k) hardware RAID controller that can touch software RAID's 
>> performance on Linux, especially in "challenging" setups, like 
>> RAID-5.  Sure, software RAID eats more CPU, but most PCs have CPU to 
>> spare these days.  Would you rather eat 10% of one of your Xeon CPUs 
>> to get 200 MB/sec or 100% of an Intel 960 to get 15 MB/sec?
>
> While this is certainly true, in the context of Asterisk you also have 
> to consider the extra PCI bus usage for all this data going back and 
> forth to the drives while the RAID parity/mirror stuff is being done.

Sure.  Particularly after this week's echo discussion.

However, there are three ways around this.  First, modern chipsets 
don't route their on-board IDE ports over a PCI bus--they're local to 
the southbridge and routed over a high-speed link between the north and 
south bridges.  Second, most of the current server boards on the market 
have 3 PCI busses--typically 2 PCI-X busses and 1 PCI 32/33 bus.  So, 
just make sure that you don't put your IDE drives onto the same bus as 
your telephony cards, and you should be fine.  One quick way to 
check--you can only have 1 PCI-X 133 slot per bus.  So, if a board 
advertises two PCI-X 133 slots, then you know that you have at least 2 
PCI-X busses.  PCI-X 100 allows 2 slots, and PCI-X 66 allows 4.

One more note: at a previous job, I actually had problems with 3ware 
6000-series cards and PCI latency.  I had 3 4-port 3ware cards sharing 
a bus with an Intel GigE card, and under heavy disk load, I was losing 
packets due to PCI latency.  Since these were primarily NFS servers, 
and NFS is hyper-sensitive to packet loss, this killed my performance.  
Turning down the PCI latency timer on the 3ware cards helped quite a 
bit.  Modern systems with multiple busses shouldn't have this problem.


Scott