Re: [boost] serialization performance

From: Robert Ramey (ramey_at_[hidden])
Date: 2005-11-12 12:39:50

Martin Slater wrote:

> Just to reiterate my experience of profiling boost::serialization
> loading under vtune (I cannot recommend this enough for serious
> performance analysis) using one of our real data files for testing
> (weighs in at > 100mb and thousands of individual instances of
> objects) that the major bottleneck was strcmp caused by the type_id
> compare looking up the per type information used for tracking and the
>  such. This outweighed everything else by a wide margin. I can dig up
> the details if your interested and up some point I will need to look
> at this again and try and optimise it. This was using 1.32.0 so take
> the above with a grain of salt *if* this area has changed
> significantly.

A couple of observations.

I believe the intel machine has hardware instructions which implement
strcmp and that compilers support them.  So even if strcmp is the
bottleneck, I wouldn't expect it to show up on the profiler unless
some sort of inlining were turned off.  Or maybe the vtune profiler
has special provision for these cases somewhere.

I did check to verify that the strcmp in the type-id lookup has been
removed.  Instead we just make sure there is only one instance
of a particular extended_type_info record so that we can
just compare the addresses.  There are still some optimizations
to be implemented - but I can't predict how much they will
speed up anything.

Robert Ramey