The json serializer's performance (when using the C speedups) appears to
be tied to the depth of the structure being serialized on some systems.
 In particular, dict structure that are more than a few levels deep,
especially when they content mixed values (lists, strings, and other
dicts) causes severe serialization penalties (in the neighborhood of an
extra 20-30 seconds) on some systems.

On SPARC systems, this is very likely because of the recursive call
structure that the C speedups serializer uses which doesn't work well
with SPARC due to register windows.

On x86 systems, recursive call structures are cheap, so this doesn't
appear to be an issue there.

SPARC systems with higher amounts of memory bandwidth don't suffer as
badly from this (such as Niagra T1000, T2000, etc. systems), but older
UltraSparc systems are especially prone to performance issues.