Re: Compression filter for Loopback device

[Paulo Marques]

On Fri, 2004-07-23 at 19:20, Phillip Lougher wrote:
> On Thu, 2004-07-23 Paulo Marques wrote:
>  >
>  >I did start working on something like that a while ago. I even
>  >registered for a project on sourceforge:
>  >
>  >http://sourceforge.net/projects/zloop/
>  >
>  >    - The block device doesn't understand anything about files. This is
>  >an advantage because it will compress the filesystem metadata
>  >transparently, but it is bad because it compresses "unused" blocks of
>  >data. This could probably be avoided with a patch I saw floating around
>  >a while ago that zero'ed delete ext2 files. Zero'ed blocks didn't accupy
>  >any space at all in my compressed scheme, only metadata (only 2 bytes
>  >per block).
>  >
> 
> The fact the block device doesn't understand anything about the 
> filesystem is a *major* disadvantage.  Cloop has a I/O and seeking 
> performance hit because it doesn't understand the filesystem, and this 
> will be far worse for write compression.  Every time a block update is 
> seen by your block layer you'll have to recompress the block, it is 
> going to be difficult to cache the block because you're below the block 
> cache (any writes you see shouldn't be cached).  If you use a larger 
> compressed block size than the block size, you'll also have to 
> decompress each compressed block to obtain the missing data to 
> recompress.  Obviously Linux I/O scheduling has a large part to play, 
> and you better hope to see bursts of writes to consecutive disk blocks.

Yes, I agree it is a major disadvantage. That is way I listed this as
one of the reasons to drop the project altogether :)

Anyway, my main concern was compression ratio, not performance. 

Seek times are very bad for live CD distros, but are not so bad for
flash or ram media. 

>  >I did a proof of concept using a nbd server. This way I could test
>  >everything in user space.
>  >
>  >With this NBD server I tested the compression ratios that my scheme
>  >could achieve, and they were much better than those achieved by cramfs,
>  >and close to tar.gz ratios. This I wasn't expecting, but it was a nice
>  >surprise :)
> 
> I'm very surprised you got ratios better than CramFS, which were close 
> to tar.gz.  Cramfs is actually quite efficient in it's use of metadata, 
> what lets cramfs down is that it compresses in units of the page size or 
> 4K blocks.  Cloop/Squashfs/tar.gz use much larger blocks which obtain 
> much better compression ratios.
> 
> What size blocks did you do your compression and/or what compression 
> algorithm did you use?  There is a dramatic performance trade-off here. 
> If you used larger than 4K blocks every time your compressing block 
> device is presented with a (probably 4K) block update, you need to 
> decompress your larger compression block, very slow.  If you used 4K 
> blocks then I cannot see how you obtained better compression than cramfs.

You are absolutely correct. I was using 32k block size, with 512 byte
"sector size". A 32k block would have to compress into an integer number
of 512 byte sectors. Most of my wasted space comes from this, but I was
assuming that this would have to work over a real block device, so I
tried as much as possible to make every read/write request to the
underlying file to be "512-byte block"-aligned.

The compression algorithm was simply the standard zlib deflate.

But has I said before, my major concern was compression ratio. 

I left the block size selectable on mk.zloop, so that I could test
several block sizes and measure compress ratio / performance. 

>From what I remember, 4k block sizes really hurt compression ratio. 32k
was almost as good as 128k or higher.

What I would really like to know is if anyone has real world
applications for a compression scheme like this, or is this just a waste
of time...

-- 
Paulo Marques - www.grupopie.com
"In a world without walls and fences who needs windows and gates?"

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/