Magma object serializer

[Chris Muller]

> I am not sure if I understood. In our benchmarks, we have a list of samples
> and each sample is at the same time an array of objects that we
> serialize/materialize. For each sample we create instantiate a serializer
> and a materializer.
>
> I understood now that we should reuse the same serializer/materialize
> instance for all samples?  if true, why I don't use a Singleton ? I
> mean...it is not clear for me when to instantiate a serializer.

Users of MaObjectSerialization only instantiate one serializer to
handle a related "groups" of objects - usually related by being of the
same set of classes.  In this case, if the samples of your benchmark
are small (e.g., < 100 obejcts), you should reuse the same serializer
for each sample being serialized/materialized.

A generic pattern for improving real-world performance is to off-load
work to an initialization step.  For example, many applications
pre-cache certain objects from a database at system startup (a.k.a.,
the "initialization step").  Since startup of the system is only done
once, it is ok if it takes, say, an additional 10 or 30 seconds to
pre-cache if it means that users will have sub-second response times
after the system comes up rather than response times of 5 seconds..

MaObjectSerialization uses this pattern - it is expensive to
instantiate a MaObjectSerializer (about 500 milliseconds) but, in
exchange, the performance of the serializer is improved.

So if the benchmark is going to include initialization of a new
MaObjectSerializer for each of _many tiny_ "samples", then that is not
a good measurement of how it would be used in actual practice.  The
repeated initializations time will dominate 99% of the time consumed,
and the "benchmark" would favor the serializers which have fast
initialization times but, in fact, may be slower for serialization
and/or materialization.

This is why I stressed it is important to measure each operation -
initialization, serialization, and materialization - individually, so
that interpretation of the results can be made with respect to how it
would be used.

It's up to you, of course.  But to bring the benchmark into a
real-world usage pattern for MaObjectSerializer, I hope you will
consider:  1) measure and report initialization, serialization and
materialization separately, 2) reuse the same serializer for all of
the tiny samples or 3) use very large samples for the benchmark, so
that the initialization cost is not such a large factor.

Regards,
  Chris