[cvsnt] Re: CVS lock server

[Tony Hoyle]

alexey-panchenko at hotmail.ru wrote:
> In the request_lock() function the whole map is scanned, very much
> like a list. With large number of concurrent locks scan time will
> increase linearly.

Luckily the numbers don't get that large normally.

> I suggest add additional "map<string, map<site_t,Lock*> >" keyed by
> uppercased path. "map<site_t,Lock*>" will contain pointers to Lock
> structures having the same uppercased path keyed by lockId. So,
> instead of the whole map scan we will scan only the locks having the
> same uppercased path, resulting in scan time O(log N).
> 
> The disadvantage is that with suggested implementation we would not
> notice leaking locks for some more time. ;-)

It really doesn't make a lot of difference - the time taken to setup the 
map and copy the extra strings around offsets the advantages.  The 
common case is maybe half a dozen locks active at a time, where the 
linear scan isn't significantly better than the lookup of a map (in the 
general case 90% of the loop is incrementing a pointer and comparing two 
  integers.  If you introduce a map it will always do string compares 
which is slower).

Tony