Re: [ANNOUNCE] New hardware - SGA155D dual STM-1/OC3 PCI ad

[Stewart Smith <stewart@linux.org.au>]

On Wed, 2003-09-10 at 00:55, Alan Cox wrote:
> > 4. Optionally - and if I have enough time - I'd like
> >    to develop a twin-linear filesystem driver for
> >    time-stamped capture/playback for multiple channels
> >    of data - like a multi-band magnetic tape.
> >    BTW do you know an existing one?
> 
> I've seen people do this in user space (just interleaving the disk in
> big chunks in the app and driving it with O_DIRECT raw access) but not
> in kernel file system space.

(from memory) I think that ext2/ext3 does (or at least did) this - they
lacked any smart logic for rapid allocations - at least for inodes in
the same cylinder group. I think this was mentioned in the "Journaling
the ext2 filesystem" paper.

This could probably be faked by taking out any intelligence in block
allocation (allocate last block+1 or some such thing). Even as a mount
option (seq_alloc), this could be useful (for this type of streaming).
This will give you great write throughput, but if you don't read things
off the same way you read them - reading is going to suck.

I read in a discussion of multimedia filesystems (for PVRs) that a block
size of 256KB helped in throughput when playback configurations weren't
known (the more data you read before seeking the better). google for
"multimedia filesystems" - you'll find a fair few papers on such things.

Things like XFS were designed for large, high bandwidth systems, so
that's also worth looking into as a zero-effort approach :)