performance issues

[lunixbochs]

I'm profiling/debugging performance bottlenecks in the python client.

Two things stuck out so far.

• Calls to BufferedFormatter in logging, when logging was disabled. This can be fixed by gating any debug(), warn() etc calls behind an if _debug statement as shown in asyncio.
  • (Edit: fixed, see below)
• walk(to_nvim) and walk(from_nvim) are extremely expensive calls on nested object trees. For example, if I have a 5000 line buffer and get/set the entire thing, these calls seem to account for at least 80% of the 20ms spent. I fixed this by adding the following functions on the Remote and Nvim classes and invoking it for buf_get_lines and buf_set_lines instead of request_raw:

class Remote:
    def request_raw(self, name, *args, **kwargs):
        return self._session.request_raw(name, self._session._to_nvim(self), *args, **kwargs)

class Nvim:
    def request_raw(self, name, *args, **kwargs):
        return self._session.request(name, *args, **kwargs)

Profiling data generated with yappi on Python 3.3 and my test script are attached.
after.txt
before.txt

import neovim

sample = ['a'] * 10000

def setup():
    n = neovim.attach('child', argv=['/Users/aegis/build/neovim/build/bin/nvim', '--embed'])
    buf = n.current.buffer
    buf[:] = sample
    return buf

def bufeq(buf):
    buf[:] = sample

def eqbuf(buf):
    _ = buf[:]

buf = setup()
for i in range(10):
    bufeq(buf)

for i in range(10):
    eqbuf(buf)

import sys
sys.exceptionhandler = lambda *a: 0

[justinmk]

ref: liuchengxu/vim-clap@75eb77e#r36434357

[justinmk]

This can be fixed by gating any debug(), warn() etc calls behind an if _debug statement as shown in asyncio.

@lunixbochs logging is now removed in the "dist" package (i.e. what we publish to pypi), try Pynvim 0.4.1. dd45f83