Le dimanche 12 juillet 2009 à 21:48 +0200, Jakub Zawadzki a écrit :
> Hi,
>
> This patch (Proof of Concept) removes allocating memory for columns data,
> and makes them 'dynamic' (packets redissected when column data needed)
>
> Getting column data is done in packet_list_record_get_column()
> I'd be grateful if someone could give some hints how it could be optimized.
> (Anyway I think it'd be nice if we have some column cache for some previous and next rows, or e.g. tcp.stream related)
>
> Unfortunetly resorting for large dumps is PITA - we could also cache some columns (configurable by user?)
>From my own test (cf http://wiki.wireshark.org/Development/Optimization)
I would say that you always want to cache columns so it's only slow the
first time but:
- you don't need to cache some of them, frame number and times.
- use a cache for addresses
- don't copy constant data (protocol and col info use a lot of them).
Memory usage is not that bad:
my patched version with a 53MB file (250k packets)
virt res
137 84 MB
after sorting
157 106 MB
svn (compiled without WANT_GUARD_PAGES and DEBUG_USE_CANARIES)
368 315 MB
but there's the copy of col_expr in frame data.
And on a bigger file (doesn't load in svn Wireshark):
557 MB 1.4M packets
virt res
470MB 420MB
after sorting
600MB 518MB
Didier
