Skip to content

Files

Latest commit

 

History

History
274 lines (195 loc) · 6.15 KB

greenlet-profiling.md

File metadata and controls

274 lines (195 loc) · 6.15 KB

Profiling Greenlets

Yappi now supports profiling greenlets!

How do you use it?

Profile simple greenlets

Basic Usage

Here's an example of profiling a simple greenlet based application which runs two greenlets

import yappi
from greenlet import greenlet
import time

## Application logic

def burn_cpu(secs):
    t0 = time.process_time()
    elapsed = 0
    while (elapsed <= secs):
        for _ in range(1000):
            pass
        elapsed = time.process_time() - t0

class GreenletA(greenlet):
    def run(self):
        burn_cpu(0.1)

class GreenletB(greenlet):
    def run(self):
        burn_cpu(0.2)

# Running the profiler:

# Step 1: Configure the profiler to work with greenlets

yappi.set_context_backend("greenlet")
yappi.set_clock_type("cpu")

# Step 2: Run the profiler and stop it

yappi.start()

a = GreenletA()
b = GreenletB()

a.switch()
b.switch()

yappi.stop()

# Step 3: View results

print("## Function stats:")
yappi.get_func_stats().print_all()

print("\n## Greenlet stats:")
yappi.get_greenlet_stats().print_all()

Sample Output:

## Function stats:

Clock type: CPU
Ordered by: totaltime, desc

name                                  ncall  tsub      ttot      tavg
test.py:7 burn_cpu                    2      0.255467  0.300078  0.150039
test.py:20 GreenletB.run              1      0.000009  0.200057  0.200057
test.py:16 GreenletA.run              1      0.000008  0.100038  0.100038

## Greenlet stats:

name           id     ttot      scnt
GreenletB      3      0.200074  1
GreenletA      2      0.100051  1
greenlet       1      0.000076  3

The meaning of each column and table is explained here - Introduction

Profile gevent applications

With support for greenlets, you can now profile popular async frameworks built on top of greenlets like Gevents.

Basic Usage

import yappi
from gevent import Greenlet
import time

## Application logic

def burn_cpu(secs):
    t0 = time.process_time()
    elapsed = 0
    while (elapsed <= secs):
        for _ in range(1000):
            pass
        elapsed = time.process_time() - t0

class GreenletA(Greenlet):
    def _run(self):
        burn_cpu(0.1)

class GreenletB(Greenlet):
    def _run(self):
        burn_cpu(0.2)

# Running the profiler:

# Step 1: Configure the profiler to work with greenlets

yappi.set_context_backend("greenlet")
yappi.set_clock_type("cpu")

# Step 2: Run the profiler and stop it

yappi.start()

a = GreenletA()
b = GreenletB()

a.start()
b.start()
a.get()
b.get()

yappi.stop()

# Step 3: View results

print("## Function stats:")
yappi.get_func_stats().print_all()

print("\n## Greenlet stats:")
yappi.get_greenlet_stats().print_all()

Sample output:

## Function stats:

Clock type: CPU
Ordered by: totaltime, desc

name                                  ncall  tsub      ttot      tavg
tests/test_random.py:7 burn_cpu       2      0.257554  0.300067  0.150033
..s/test_random.py:20 GreenletB._run  1      0.000007  0.200025  0.200025
..s/test_random.py:16 GreenletA._run  1      0.000009  0.100058  0.100058
..
.. More function stats
..

## Greenlet stats:

name           id     ttot      scnt
GreenletB      4      0.200048  1
GreenletA      3      0.100075  1
greenlet       1      0.006496  2
Hub            2      0.000212  1

With 'threading' monkey patched

When the threading module is monkey patched, threading.Thread is used to spawn greenlets instead of gevent.Greenlet. Since yappi reports the name of each greenlet class by default, the user must inform yappi to retrieve the class name from the threading library instead. Yappi provides set_context_name_callback to do so. See below for an example:

from gevent import monkey
monkey.patch_all()

import yappi
import threading
import gevent
import time

## Application logic

def burn_cpu(secs):
    t0 = time.process_time()
    elapsed = 0
    while (elapsed <= secs):
        for _ in range(1000):
            pass
        elapsed = time.process_time() - t0

class ThreadA(threading.Thread):
    def run(self):
        burn_cpu(0.1)

class ThreadB(threading.Thread):
    def run(self):
        burn_cpu(0.2)

# Running the profiler:

# Step 1: Configure the profiler to work with greenlets
yappi.set_context_backend("greenlet")
yappi.set_clock_type("cpu")

# Step 2: Configure the system to capture thread names correctly
def _ctx_name_callback():
    curr_gl = gevent.getcurrent()
    if curr_gl is gevent.get_hub():
        return curr_gl.__class__.__name__
    # yappi._ctx_name_callback returns the name of the thread
    # class
    return yappi._ctx_name_callback()

yappi.set_context_name_callback(_ctx_name_callback)

# Step 3: Run the profiler and stop it
yappi.start()

a = ThreadA()
b = ThreadB()

a.start()
b.start()
a.join()
b.join()

yappi.stop()

# Step 4: View results

print("## Function stats:")
yappi.get_func_stats().print_all()

print("\n## Greenlet stats:")
yappi.get_greenlet_stats().print_all()

Sample Output:

## Function stats:

Clock type: CPU
Ordered by: totaltime, desc

name                                  ncall  tsub      ttot      tavg
tests/test_random.py:11 burn_cpu      2      0.255339  0.300063  0.150031
..hreading.py:870 ThreadB._bootstrap  1      0.000011  0.200229  0.200229
..ng.py:901 ThreadB._bootstrap_inner  1      0.000059  0.200218  0.200218
tests/test_random.py:24 ThreadB.run   1      0.000010  0.200042  0.200042
..hreading.py:870 ThreadA._bootstrap  1      0.000012  0.100218  0.100218
..
.. More function stats
..

## Greenlet stats:

name           id     ttot      scnt
ThreadB        4      0.200243  1
ThreadA        3      0.100228  1
_MainThread    1      0.000875  4
Hub            2      0.000182  3

Limitations in gevent profiling

Gevent allows users to run functions on a pool of native threads via ThreadPool. All threads spawned as part of this pool cannot be tracked by yappi and so yappi cannot report stats for functions / greenlets running on them.