Andrew Grillet wrote in a message to Mike Bilow:

 AG> I had a discussion with someone in which it was suggested it
 AG> might be possible to compress stuff being swapped. Not to
 AG> save disk space, which is a minor issue, but to reduce
 AG> amount of data transfer, as it is slow.

 MB> You could do it, and IBM has looked at the idea for OS/2.  However,
 MB> the end result of the analysis is that disk space is cheap and
 MB> plentiful, while CPU time for compression and decompression is
 MB> neither.

 AG> My point was that DISK TRANSFER TIME is very expensive. If
 AG> you run an IDE based system, the CPU does the transfers - if
 AG> it does compression first and the transfer is twice as fast,
 AG> then all that is needed is that the compression takes half
 AG> the time a disk-write takes - not a great achievement! 

 AG> Decompression is normally faster than compression,

Don't forget that you need to compare the physical transfer time of the
uncompressed data against the compression time plus the transfer time of
the compressed data.  In all probability, simply transferring uncompressed
data is going to be faster.  Given that you must run compression over 4 KB
pages, or you cannot swap them, the efficiency of compression is going to
suffer.  Even the processing overhead will be substantial to compress 4 KB
blocks, with loading and unloading dictionaries.

 MB> Therefore, compressing the swapfile is just not economical.

 AG> Very true on my SCSI based server, but not on the IDE based
 AG> workstations. 

IDE systems use different access methods.  Some use PIO, as you say, but
others use DMA.  Especially with local bus and PCI IDE, DMA is more common.
 
-- Mike


---