[vdr] Re: Kernel-Patch to make VDR undisturbable

[Markus Dahlweid <markus@dahlweid.com>]

Hi Andreas!

Andreas Schultz wrote:

>This is all very interresting and it will improve things, but IMHO does *NOT* 
>really help us. Why? Because all the mentioned patches deal with scheduling 
>latency. They are designed to give you scheduling latencies < 3ms and we 
>don't really need that.
>In the DVB datastream, audio samples are usally around 32ms and a single 
>frames is arround 40ms. The 1MB buffer we use, appears to hold about 500ms 
>worth of audio and video data. The dvb cards buffer apparently also hole 
>something of arroung 100ms (just a guess) of data. A scheduling latency of 
>lets say 100ms will therefore hardly affect us. Normal rt scheduling on a 
>stock linux kernel will give us latencies that are somewhat smaller than 
>that.
>What we would need is a guaranteed disk I/O rate, and this is something i 
>haven't found yet.
>
>Yet, i might be wrong, so i'm going to try them anyway.....
>
I agree. 500ms of buffer should normally be sufficient for the DVB cards 
data. The problem are those few times, where the Linux kernel cannot 
guarantee a response within this time limit. The low latency patch and 
the preemptive patch ensure only, that the latency between an interrupt 
and the corresponding process beeing scheduled, is reduced to a few ms. 
Another problem is the flushing of the CPU caches. In a students project 
at University Bremen some people found out, that a complete flush of the 
CPU second level cache can take more than 500ms (depending on the 
processor). AFAIK there is no possibility to prevent the complete cache 
flush, if there are unfavourable situations, which makes the kernel to 
so. Normally the user does not reckon those situations, but if both 
situaltions occur simultaniously, data of the card may be lost.

I am very well aware, that those situations are only the worst case 
scenario, which may only happen once a week under heavy load, but they 
may eventually happen. In my opinion all patches, which ensure a "more 
real-time" behaviour of the kernel, should be used to get those VDR 
boxes even more stable.

... just my two cents.

Markus