Issue 32433: Provide optimized HMAC digest (original) (raw)

hmac.HMAC's flexibility comes with a cost. To create the MAC of a message, it has to create between three and four hash instances (inner, outer, key for long keys, result), multiple bound method objects and other temporary objects. For short messages, memory allocation, object handling and GIL release/acquire operations dominate the performance.

I propose to provide a fast one-shot HMAC function: hmac.digest(key, msg, digstmod) -> bytes function. A PoC implementation based on OpenSSL's HMAC() function and a pure Python implementation as inlined HMAC showed promising performance improvements. The C implementation is 3 times faster.

Standard HMAC: $ ./python -m timeit -n200000 -s "import hmac" -- "hmac.HMAC(b'key', b'message', 'sha256').digest()" 200000 loops, best of 5: 5.38 usec per loop

Optimized Python code: $ ./python -m timeit -n 200000 -s "import hmac; hmac._hashopenssl = None" -- "hmac.digest(b'key', b'message', 'sha256')" 200000 loops, best of 5: 3.87 usec per loop

OpenSSL HMAC() $ ./python -m timeit -n 200000 -s "import hmac" -- "hmac.digest(b'key', b'message', 'sha256')" 200000 loops, best of 5: 1.82 usec per loop