nnrpd, cnfs and binaries

[Alex Kiernan]

Russ Allbery <rra at stanford.edu> writes:

> bill davidsen <davidsen at tmr.com> writes:
> 
> > Typically the startup is much faster with -D and the memory used is less
> > due to the data being shared COW for most functional VM implementations.
> > Obviously YMMV.
> 
> I don't think -D helps that much there because most of the memory usage is
> from opening the storage and overview backends, and that's done after the
> fork even in -D mode.
> 

If you're just using "safe" backends (CNFS and tradindexed), you can
hoist the SetupDaemon call (which is what we've done); without this we
really struggle to keep up with the page mapping which happens at
startup (large CNFS bitmaps being the main area).

The other thing we have (which I still haven't found time to commit)
is pre-forking which helps lots - we have a big peak at 6pm (the UK's
traditional cheap rate call time) when we see high connect rates after
which it takes a couple of hours to repopulate the server pool.

-- 
Alex Kiernan, Principal Engineer, Development, THUS plc