Re: 2.6.x: iowait problem while burning a CD

[Bill Davidsen <davidsen@tmr.com>]

Rik van Riel wrote:
> On Fri, 27 Feb 2004, Paolo Ornati wrote:
> 
> 
>>trying to burn a CD "on the fly" I have noticed a strange thing. During the 
>>burning "iowait" remains enough low (~3%, MAX 10%) but, after a little 
>>time, it suddenly and quickly goes up to 80-90%: in this condition MKFS 
>>seems unable to fill the FIFO buffer as quickly as the CD-writer writes  
> 
> 
>>Any ideas?
> 
> 
> At that point, mkisofs is probably running into a bazillion
> small files, in subdirectories all over the place.
> 
> Because a disk seek + track read takes 10ms, it's simply not
> possible to read more than maybe 100 of these small files a
> second, so mkisofs can't keep up.

I certainly have seen this, although there is something odd in the 
process with large files as well. I sometimes see it in 2.4 as well, so 
it may be a characteristic of the application.

What to do about it (things to try, not requirements):

- use fs= to increase the size of the fifo a bit, remember to put "m"
   after the buffer size or it will be taken in bytes. This helps with
   the problem Rik describes
- be sure DMA is on the CD and the drive
- allow ints during io
- I'm told deadline schedular but haven't tried it
- get a new version of cdrecord, build with the 2.6 headers, use
   ATAPI:/dev/hdX for dev= instead of ide-scsi (this is probably not
   critical for data CD in 2.6.2 and later, ide-scsi seems to work)
- use burn-free
- drop the speed of the burn
- build the ISO image first, burn from that instead of a pipe
- use growisofs for DVD

-- 
bill davidsen <davidsen@tmr.com>
   CTO TMR Associates, Inc
   Doing interesting things with small computers since 1979