Python 3.11 performance with frame pointers

Hi,

As part of the proposal to enable frame pointers by default in Fedora
(Changes/fno-omit-frame-pointer - Fedora Project Wiki), we
did some benchmarking to figure out the expected performance impact.
The performance impact was generally minimal, except for the
pyperformance benchmark suite where we noticed a more substantial
difference between a system built with frame pointers and a system
built without frame pointers. The results can be found here:
GitHub - DaanDeMeyer/fpbench (look at the mean difference
column for the pyperformance results where the percentage is the
slowdown compared to a system built without frame pointers). One of
the biggest slowdowns was on the scimark_sparse_mat_mult benchmark
which slowed down 9.5% when the system (including python) was built
with frame pointers. Note that these benchmarks were run against
Python 3.11 on a Fedora 37 x86_64 system (one built with frame
pointers, another built without frame pointers). The system used to
run the benchmarks was an Amazon EC2 machine.

We did look a bit into the reasons behind this slowdown. I’ll quote
the investigation by Andrii on the Fesco issue thread here
(https://pagure.io/fesco/issue/2817):

So I did look a bit at Python with and without frame pointers trying to
understand pyperformance > regressions.

First, perf data suggests that big chunk of CPU is spent in _PyEval_EvalFrameDefault,
so I looked specifically into it (also we had to use DWARF mode for perf for apples-to-apples
comparison, and a bunch of stack traces weren’t symbolized properly, which just again
reminds why having frame pointers is important).

perf annotation of _PyEval_EvalFrameDefault didn’t show any obvious hot spots, the work
seemed to be distributed pretty similarly with or without frame pointers. Also scrolling through
_PyEval_EvalFrameDefault disassembly also showed that instruction patterns between fp
and no-fp versions are very similar.

But just a few interesting observations.

The size of _PyEval_EvalFrameDefault function specifically (and all the other functions didn’t
change much in that regard) increased very significantly from 46104 to 53592 bytes, which is a
considerable 15% increase. Looking deeper, I believe it’s all due to more stack spills and
reloads due to one less register available to keep local variables in registers instead of on the stack.

Looking at _PyEval_EvalFrameDefault C code, it is a humongous one function with gigantic switch
statement that implements Python instruction handling logic. So the function itself is big and it has
a lot of local state in different branches, which to me explained why there is so much stack spill/load.

Grepping for instruction of the form mov -0xf0(%rbp),%rcx or mov 0x50(%rsp),%r10 (and their reverse
variants), I see that there is a substantial amount of stack spill/load in _PyEval_EvalFrameDefault
disassembly already in default no frame pointer variant (1870 out of 11181 total instructions in that
function, 16.7%), and it just increases further in frame pointer version (2341 out of 11733 instructions, 20%).

One more interesting observation. With no frame pointers, GCC generates stack accesses using %rsp
with small positive offsets, which results in pretty compact binary instruction representation, e.g.:

0x00000000001cce40 <+44160>: 4c 8b 54 24 50 mov 0x50(%rsp),%r10

This uses 5 bytes. But if frame pointers are enabled, GCC switches to using %rbp-relative offsets,
which are all negative. And that seems to result in much bigger instructions, taking now 7 bytes instead of 5:

0x00000000001d3969 <+53065>: 48 8b 8d 10 ff ff ff mov -0xf0(%rbp),%rcx

I found it pretty interesting. I’d imagine GCC should be capable to keep using %rsp addressing just fine
regardless of %rbp and save on instruction sizes, but apparently it doesn’t. Not sure why. But this instruction
increase, coupled with increase of number of spills/reloads, actually explains huge increase in byte size of
_PyEval_EvalFrameDefault: (2341 - 1870) * 7 + 1870 * 2 = 7037 (2 extra bytes for existing 1870 instructions
that were switched from %rsp+positive offset to %rbp + negative offset, plus 7 bytes for each of new 471 instructions).
I’m no compiler expert, but it would be nice for someone from GCC community to check this as well (please CC
relevant folks, if you know them).

In summary, to put it bluntly, there is just more work to do for CPU saving/restoring state to/from stack. But I don’t
think _PyEval_EvalFrameDefault example is typical of how application code is written, nor is it, generally speaking,
a good idea to do so much within single gigantic function. So I believe it’s more of an outlier than a typical case.

We have a few questions:

  • Is this slowdown when Python is built with frame pointers to be
    expected? Has the Python community done any of their own experiments
    with building Python with and without frame pointers?
  • Is there anything we can do to fix the slowdown when Python is built
    with frame pointers?
  • Should we expect any change in benchmark results if we benchmark
    against Python 3.12? Supposedly there are changes in Python 3.12
    related to frame pointers so we’re wondering if those changes might
    affect these results in any way.

Cheers,

Daan De Meyer

4 Likes

There’s already work to look at making managing the eval loop code easier, e.g. Mechanics of interpreter generation from DSL · Issue #479 · faster-cpython/ideas · GitHub . That will make potentially addressing it’s size, etc. easier.

2 Likes

There’s another thread about this started by someone from Fedora.

As @brettcannon mentioned, we’re working on generating the evaluation loop from a DSL. This will allow us to make larger changes to how the instructions are evaluated, and hopefully fix this regression.

For the time being, we could potentially use __attribute__((optimize("omit-frame-pointer")) in the eval function so that, when distros build Python, they don’t see the 10% regression. Of course, we wouldn’t have the benefits of a frame pointer in that function, but it beats disabling it for everything.

1 Like

I suspect that this is not useful to do.
Fedora is already going to opt python out of the frame pointer change.
And its very compiler/cpu-architecture specific.

Looking forward to seeing what the DSL generation creates.