There is already a parameter for memoize:
ns_cache create util_memoize -size
[ad_parameter -package_id [ad_acs_kernel_id] MaxSize memoize 200000]
I think the default of 200k bytes is absurdly small as a default for
a production machine.
More memory is taken by the nsv's and of course
they cannot really be sized. I checked and here are the
top ten nsv's by size
entries bytes nsv
170 5236 locale
49 8811 site_nodes
41 10522 ad_page_contract_filters
192 54013 api_library_doc
1580 97454 apm_library_mtime
2190 170279 proc_source_file
2189 336690 proc_doc
94 574815 apm_version_properties
2189 1115145 api_proc_doc
3134 1332667 OACS_FULLQUERIES
The memory footprint of my server is 37mb and as far as I can tell,
there is only ~4mb cached (although there are some nsv's I skipped
like the parameter cache and and some others but I think they
are smaller ones). This is on my testing site with all packages
installed and mounted (some several times).
Anyway, the short answer is that things are bloated and it
does not seem to really be overly aggressive caching.
I think the real issue is that there are just a huge number of
procedures defined.
The source for all the tcl libraries is ~3mb and I think when they get bytecode compiled you end up keeping src and bytecode around so that
is a big piece of the pie. Also, I am not sure how much compiled
templates end up taking, although they are done on demand so an empty
idle server should not have very many loaded.
I suppose we could look at making query cached on demand, put
in a size limited ns_cache, or not cached at all. Also we could ad a
parameter to disable proc doc stuff by default and that might help but
again, it oes not seem like either of these would make a huge difference.
