Hard disk speed - video system - OT - Mainly for Will
Here is some more info on what has been going on with the surveillence
project. - Related to my previous posts.
The number that I was looking at with procinfo (IOwait) is now hovering
around 55%. This is about 10% higher than it was before the RAM
doubling. There have however been too many changes to the way the system
works to put any credence in even suggesting that was the cause! I have
also been spending a lot of fun time looking at the disk IO schedular in
kernel 2.6. Basically you can change the buffer sizes and waits by
writing to sysfs and in newer kernels you can even change the queueing
method on the fly. You have to dig to find good articles on it and it is
worth looking at.
I am in the midst of doing some I/O rate tests but I dont have all the
hard facts lined up yet. At the sake of boring you here is some of the
With a design rate of 5FS and 6 cameras of about 100KByte images each
the disk writing rate in capture mode is about 3 Megabytes per second.
The individual hard drive specs say it can transfer around 72 Megabytes
per second. This is also borne out by hdparm -tT. With RAID0 disks that
would be even faster.
In real life and with no other system load I am actually getting about
3-4FPS from one camera type and 1-2 from another. I have some more work
to do in this area but whatever way you look at it the write rate is
probably 2/3 of design.
With "system load" I can drop the capture rate to less than 1FPS. What
load that is I am still collating tests on..
I suddenly dawned on me that since I have a mix of 54MB Wireless and
100MB Ethernet I am getting mighty close to saturating the achievable
datarate over those connections. 54MB is of course aggregate bandwidth
but we also have a front desk PC reading from the network server over
the same link that 3 of the cameras send on. It's only about 200
kytes/second but it does contribute. Back at the server some other
cameras, PC's and now the WAP also talk to the capture server. I use to
use 30% bandwidth for sizing layer 1 stuff and maybe 50% for layer 2
(switched) but I think that with my NFS over the wire processing and
other I/O is not disk I/O or CPU bound, its the Ethernet connection! I
might setup a separate routed network and put two NICs in the capture box.
Do you agree or disgaree with the maths/deduction? I might also dump the
output from ifconfig to see exactly what kind of NIC throughout is going
To be continued...