[mythtv-users] pdflush is killing me

Dan Ritter dsr-myth at tao.merseine.nu
Wed Feb 7 17:51:15 UTC 2007

On Wed, Feb 07, 2007 at 10:08:44AM -0500, Steven Adeff wrote:
> gotcha. So in our case, where large amounts of writes are occuring
> over a long period of time (like recording 3 one hour HDTV shows), we
> want don't need this caching, and waiting to long may make writing in
> bursts actually slower than writing in a constant stream?

Experimental evidence is not in.

My guess (and I am not a kernel hacker, just an experienced
sysadmin) follows...

An ATSC stream at 20Mb/s represents about 1/24th of a PATA disk's
maximum sequential write capability (numbers from a recent Boston
Linux User's Group discussion list posting on WD2500JB and S
drives). Inner-track sequential writes will max at around half that,
so about 1/12th capability. 

But we are in a multiprocessing environment, and a typical job
will be interrupted by, let's say, a request for a directory
scan and then a file read, every few tenths of a second. 

You need a sufficient cache to account for these other jobs running
concurrently; these will be influenced by seek times (9 ms average
track-to-track, so worst-case 18ms seeks, and easily 8 seeks and then a
read). Let's say 250ms of disk activity before the OS gets back to your
job. That's almost a megabyte of dirty data that you'll have accumulated.

Experimentation is desirable. Multiple spindles are desirable...


_.. ___ . ...   _ .... .   _. ... ._   ._. . ._ _..   _.__ ___ .._ ._.
__ ._ .. ._.. ..__..   _ .... .   .._. _... .. ..__..
http://tao.merseine.nu/~dsr/eula.html is hereby incorporated by reference.

More information about the mythtv-users mailing list