On Sat, 15 Jun 1996 02:43:49 -0400 (EDT), Kevin M Bealer
<kmb203@psu.edu> said:
> On Fri, 14 Jun 1996, Stephen Tweedie wrote:
>> Planned for 2.1. There are a number of places where I want to
>> improve swapping throughput further, especially to do with even better
>> clustering of writeouts and managing sequential access of large
>> datasets more intelligently. NFS swapping should also appear at some
>> point.
> This looks like it should know what the underlying device is... On IDE you
> could lose number crunching power if the kernel was babysitting the
> controller. SCSI of course would be less intrusive.
Paging throughput is already critical to performance. The whole point
about the paging algorithms is to try to reduce the amount of disk IO
being done, since it is so much slower than memory access. ALL paging
incurs a performance hit --- more so on IDE than on SCSI, as you say
--- so regardless of the underlying device we should be trying to page
as little as possible.
There IS one area where knowing the underlying device can be genuinely
useful, though. If we are paging from binaries on an NFS mounted
partition, then it can be helpful to swap those binaries to a local
fast disk rather than to keep paging them from the network. This
provides a limited page cache on local disks for NFS binaries, and can
improve performance on such systems. We would not be changing NFS
semantics here, since it is already implicit in NFS paging that pages
can be cached locally indefinitely, as long as they are referenced by
any running process.
Cheers,
Stephen.
-- Stephen Tweedie <sct@dcs.ed.ac.uk> Department of Computer Science, Edinburgh University, Scotland.