[mythtv-users] Linux software raid question
jan.ceuleers at computer.org
Thu Jun 5 18:50:01 UTC 2008
David Brodbeck wrote:
> This is a good point for small systems. One caveat is that, while the
> CPU load is low, software RAID can run into bus contention problems for
> really big arrays. Imagine you have a 10-disk array. When you write to
> it, individual commands have to go across your PCI bus to 10 different
> disks. With a hardware controller, the write goes out on the bus once,
> and the rest happens internally.
Once again, absolutely right.
My server has a really low-end motherboard (with a 1.5GHz Via C7 CPU).
The performance of my 4-drive RAID5 array, as measured by hdparm, is
limited by the PCI bus bandwidth, not by the CPU, the SATA link speed
(1.5 Gbit/s) or the disks (udma5).
I picked the low-end mobo for reasons of power consumption and heat, and
Another point though: some mobos have multiple PCI buses which don't
always contend for the same bandwidth. In my case, my 4-drive RAID5
array is served by a 4-way SATA-PCI adapter. My other 2 drives are
served by a SATA adapter that is integrated on the mobo, and that sits
directly on the combined north/southbridge.
This way the bottleneck becomes memory bandwidth, once again not the CPU.
You do really have to look at the mobo architecture, because other mobos
are structured in such a way that integrated peripherals share bus
bandwidth with plug-in cards (including another-one of my mini-ITX
More information about the mythtv-users