File system performance

[Piercarlo Grandi]

Note that foloowups have been redirected to comp.unix.internals, because
this is a discussion not limited to the 386 case.

On 1 Nov 90 11:51:48 GMT, larry at nstar.uucp (Larry Snyder) said:

larry> rcd at ico.isc.com (Dick Dunn) writes:

rcd> just that the code is cleaner and spends a little less time
rcd> fiddling with special cases that turn out not to matter.  I suspect
rcd> that cylinder groups (present in ufs but not FFS) don't help much;
rcd> they may actually be the hindrance that makes the performance
rcd> difference.

Well, actually the purpose of cylinder groups is to spread out the
ilist, and to put related inodes together and nearer the file blocks
rooted on them, not to improve IO rates for files; it does not help
bandwidth or file-based applications, but it does help interactive
performance, i.e. applications that are directory based, and open lots
of files with little IO (e.g. 'file *') or of inodes (e.g. 'ls -F').

The real problem with the BSD FFS is that the block/fragment scheme is
equivalent to allocating runs of blocks and doing synchronous 8 block
read ahead all the time, only it limits the length of the runs to 8, and
forces all reads, even random reads, to do 8 blocks of synchronous read
ahead. The authors simply did not understand the design principles of
the Unix V7 filesystem, and its author's well motivated reluctance to
souble the block size to 1024 bytes, and their use of asynchronous read
ahead instead.

Another win would have been to put the first page of the file just after
the inode.

larry> The Intel release 4 manuals go to great details to explain the
larry> gap/cylinder paramters and their "sample setting" whening making
larry> filesystems - but their examples are basically for 3b2 drives
larry> (all the way up to a big 72 meg drive).

Note that the gap should be dependent on the *CPU* and ther kernel
speed, not on the drive (for a given rotational speed). The filesystem
gap is there to compensate for interrupt and processing time.
--
Piercarlo "Peter" Grandi           | ARPA: pcg%uk.ac.aber.cs at nsfnet-relay.ac.uk
Dept of CS, UCW Aberystwyth        | UUCP: ...!mcsun!ukc!aber-cs!pcg
Penglais, Aberystwyth SY23 3BZ, UK | INET: pcg at cs.aber.ac.uk