Here are results of the benchmark which measures dump and load time for all pyc files in the stdlib (including tests).

https://bitbucket.org/storchaka/cpython-stuff/src/default/marshal/marshalbench.py

$ find * -name __pycache__ -exec rm -rf '{}' +
$ ./python -m compileall -qq -r 99 Lib
$ find Lib -name '*.pyc' | xargs ./python marshalbench.py

Without the patch:

ver.  dumps   loads      size
              770.3   19941.2
0     695.7  1178.4   19941.2
1     680.4  1180.9   19941.2
2     635.9  1115.9   19941.2
3     896.3   757.3   19941.2
4     748.0   700.6   19941.2

With marshal3_numbers.patch:

ver.  dumps   loads      size
              750.5   19946.1
0     690.2  1173.7   19946.1
1     678.2  1166.6   19946.1
2     630.9  1105.2   19946.1
3     873.2   751.5   19946.1
4     724.6   687.6   19946.1

With marshal_refs_by_value.patch:

ver.  dumps   loads      size
              780.2   19935.8
0     712.7  1202.6   19935.8
1     713.8  1195.1   19935.8
2     676.5  1126.0   19935.8
3     686.1   762.7   19935.8
4     515.1   697.6   19935.8