Directory slow down ...

[shenson@nyx10.cs.du.edu (Stephen Henson)]

Most of the points about directories have been covered already. If the directory
is too big to fit entirely in the syscache it will thrash the filesystem a lot.
This is exacerbated by the way the mintlib and some file utilities have to access
the same file several times in a row requiring a repeated lookup. A quick solution
is to increase the syscache size, either by recompiling or editing the binary.
A better way would be to add a cookie cache to the kernel relpath2cookie function,
something very simple like a single entry for the last lookup would cause a
big performance win, though a larger cache would be even better.
          I could also make the Minixfs lookup function speed up by having it
search directories from the point the last search suceeded (if at all) instead
of the start. Actually I'm looking into this for pl11.
          I did some directory access speed checks some time ago. Basically
making a directory with about 1024 entries and seeing how long it took to
search it. I found it to be rather slow, as I've implied above. Well I thought
it was slow until I tried it on a TOS filesystem ... It kept my computer busy
for a few hours if nothing else. 
          Now before anyone asks how you edit the binary here's how to do it.
1. Use the GCC utility 'cnm' to dump the symbol table of minix.xfs .
2. Note the values of the symbols _scache_size, _ucache_size and _icache_size.
3. Use your favourite binary editor to edit the binary at the points
   _?cache_size+0x1c (0x1c is the size of the binary header), these are all
   16 bits and show the cache sizes in K.

This is basically how you do binary configuration of a program, this can
all be handled automatically by a program and nlist() used to lookup the symbol
table. I never got round to finishing off my binary config program which would
do all this more easily.
          If you look at minixfs/global.h you'll see lots of other things can
be configured like this, e.g. cache mode, sync time, the 'physical partition'
options, default translation modes etc etc.
          I could say a lot more about this, it is a very useful technique.
For example you can change the default open modes (to use binary for example)
or set up default paths for things like gcc etc. It is rather cleaner than
having to override things all the time by filling up the environment.
Steve.