Jormungandr performance issue

[kinnou02]

I create this issue for speaking about the problems of performance in jormungandr. One of the slowest part in jormungandr is the marshalling phase that convert protocol buffer response in dict before serializing them in json. Another slow part that is already resolved is the deserialization of protobuf, this was solve by using the cpp implementation.

I tested a few solution:

• Pypy is another python interpreter using a JIT for better performance.
• Cython compile python code and give us the ability to type some variable, this can greatly increase performance. In this test I only active cython compilation. In almost every test this was equivalent or slower, so I will not include it in the table bellow.
• Serpy is another marshaller, it should be faster than flask_restful marshaller (that has been deprecated in recent release)

These tests were run with ab, 100 requests were made each time and the table provide the average response time.

	journey	stop_schedule	lines	lines?count=500
cpython	159	155	66	1679
pypy	172	117	55	954
cpython-serpy	NC	NC	44	1075
pypy-serpy	NC	NC	58	867
cpython-serpy-protocpp	NC	NC	27	529
pypy-serpy-protocpp	NA	NA	NA	NA
cpython-protocpp	136	102	45	1082
pypy-protocpp	NA	NA	NA	NA

protobuf doesn't support cpp optimization with pypy currently. Pypy is also a little slower on journey, it's possible that it"s because we use numpy, I will check that latter.

The migration process from marshal to serialize is slow, it is needed to rewrite the serializer for every type.

[pbougue]

Nice! Could be nice to have the original figures also (current "prod-context" times)

[kinnou02]

production is "cpython-protocpp"