Think Your Python Code Is Slow? Stop Guessing and Start Measuring

I used to be engaged on a script the opposite day, and it was driving me nuts. It labored, positive, nevertheless it was simply… sluggish. Actually sluggish. I had that feeling that this could possibly be a lot sooner if I may determine the place the hold-up was.

My first thought was to begin tweaking issues. I may optimise the information loading. Or rewrite that for loop? However I ended myself. I’ve fallen into that lure earlier than, spending hours “optimising” a bit of code solely to seek out it made barely any distinction to the general runtime. Donald Knuth had some extent when he stated, “Untimely optimisation is the foundation of all evil.”

I made a decision to take a extra methodical method. As a substitute of guessing, I used to be going to seek out out for positive. I wanted to profile the code to acquire arduous knowledge on precisely which features had been consuming the vast majority of the clock cycles.

On this article, I’ll stroll you thru the precise course of I used. We’ll take a intentionally sluggish Python script and use two improbable instruments to pinpoint its bottlenecks with surgical precision.

The primary of those instruments is known as cProfile, a strong profiler constructed into Python. The opposite is known as snakeviz, a sensible software that transforms the profiler’s output into an interactive visible map.

Establishing a improvement surroundings

Earlier than we begin coding, let’s arrange our improvement surroundings. The very best apply is to create a separate Python surroundings the place you’ll be able to set up any mandatory software program and experiment, realizing that something you do received’t affect the remainder of your system. I’ll be utilizing conda for this, however you should utilize any methodology with which you’re acquainted.

#create our take a look at surroundings
conda create -n profiling_lab python=3.11 -y

# Now activate it
conda activate profiling_lab

Now that we now have the environment arrange, we have to set up snakeviz for our visualisations and numpy for the instance script. cProfile is already included with Python, so there’s nothing extra to do there. As we’ll be working our scripts with a Jupyter Pocket book, we’ll additionally set up that.

# Set up our visualization software and numpy
pip set up snakeviz numpy jupyter

Now sort in jupyter pocket book into your command immediate. You must see a jupyter pocket book open in your browser. If that doesn’t occur robotically, you’ll doubtless see a screenful of knowledge after the jupyter pocket book command. Close to the underside of that, there might be a URL that you need to copy and paste into your browser to provoke the Jupyter Pocket book.

Your URL might be completely different to mine, nevertheless it ought to look one thing like this:-

http://127.0.0.1:8888/tree?token=3b9f7bd07b6966b41b68e2350721b2d0b6f388d248cc69da

With our instruments prepared, it’s time to have a look at the code we’re going to repair.

Our “Drawback” Script

To correctly take a look at our profiling instruments, we want a script that displays clear efficiency points. I’ve written a easy program that simulates processing issues with reminiscence, iteration and CPU cycles, making it an ideal candidate for our investigation.

# run_all_systems.py
import time
import math

# ===================================================================
CPU_ITERATIONS = 34552942
STRING_ITERATIONS = 46658100
LOOP_ITERATIONS = 171796964
# ===================================================================

# --- Process 1: A Calibrated CPU-Sure Bottleneck ---
def cpu_heavy_task(iterations):
    print("  -> Operating CPU-bound job...")
    outcome = 0
    for i in vary(iterations):
        outcome += math.sin(i) * math.cos(i) + math.sqrt(i)
    return outcome

# --- Process 2: A Calibrated Reminiscence/String Bottleneck ---
def memory_heavy_string_task(iterations):
    print("  -> Operating Reminiscence/String-bound job...")
    report = ""
    chunk = "report_item_abcdefg_123456789_"
    for i in vary(iterations):
        report += f"|{chunk}{i}"
    return report

# --- Process 3: A Calibrated "Thousand Cuts" Iteration Bottleneck ---
def simulate_tiny_op(n):
    move

def iteration_heavy_task(iterations):
    print("  -> Operating Iteration-bound job...")
    for i in vary(iterations):
        simulate_tiny_op(i)
    return "OK"

# --- Predominant Orchestrator ---
def run_all_systems():
    print("--- Beginning FINAL SLOW Balanced Showcase ---")
    
    cpu_result = cpu_heavy_task(iterations=CPU_ITERATIONS)
    string_result = memory_heavy_string_task(iterations=STRING_ITERATIONS)
    iteration_result = iteration_heavy_task(iterations=LOOP_ITERATIONS)

    print("--- FINAL SLOW Balanced Showcase Completed ---")

Step 1: Accumulating the Knowledge with cProfile

Our first software, cProfile, is a deterministic profiler constructed into Python. We will run it from code to execute our script and document detailed statistics about each operate name. 

import cProfile, pstats, io

pr = cProfile.Profile()
pr.allow()

# Run the operate you need to profile
run_all_systems()

pr.disable()

# Dump stats to a string and print the highest 10 by cumulative time
s = io.StringIO()
ps = pstats.Stats(pr, stream=s).sort_stats("cumtime")
ps.print_stats(10)
print(s.getvalue())

Right here is the output.

--- Beginning FINAL SLOW Balanced Showcase ---
  -> Operating CPU-bound job...
  -> Operating Reminiscence/String-bound job...
  -> Operating Iteration-bound job...
--- FINAL SLOW Balanced Showcase Completed ---
         275455984 operate calls in 30.497 seconds

   Ordered by: cumulative time
   Listing decreased from 47 to 10 resulting from restriction <10>

   ncalls  tottime  percall  cumtime  percall filename:lineno(operate)
        2    0.000    0.000   30.520   15.260 /dwelling/tom/.native/lib/python3.10/site-packages/IPython/core/interactiveshell.py:3541(run_code)
        2    0.000    0.000   30.520   15.260 {built-in methodology builtins.exec}
        1    0.000    0.000   30.497   30.497 /tmp/ipykernel_173802/1743829582.py:41(run_all_systems)
        1    9.652    9.652   14.394   14.394 /tmp/ipykernel_173802/1743829582.py:34(iteration_heavy_task)
        1    7.232    7.232   12.211   12.211 /tmp/ipykernel_173802/1743829582.py:14(cpu_heavy_task)
171796964    4.742    0.000    4.742    0.000 /tmp/ipykernel_173802/1743829582.py:31(simulate_tiny_op)
        1    3.891    3.891    3.892    3.892 /tmp/ipykernel_173802/1743829582.py:22(memory_heavy_string_task)
 34552942    1.888    0.000    1.888    0.000 {built-in methodology math.sin}
 34552942    1.820    0.000    1.820    0.000 {built-in methodology math.cos}
 34552942    1.271    0.000    1.271    0.000 {built-in methodology math.sqrt}

We have now a bunch of numbers that may be troublesome to interpret. That is the place snakeviz comes into its personal. 

Step 2: Visualising the bottleneck with snakeviz

That is the place the magic occurs. Snakeviz takes the output of our profiling file and converts it into an interactive, browser-based chart, making it simpler to seek out bottlenecks.

So let’s use that software to visualise what we now have. As I’m utilizing a Jupyter Pocket book, we have to load it first.

%load_ext snakeviz

And we run it like this.

%%snakeviz
important()

The output is available in two elements. First is a visualisation like this.

What you see is a top-down “icicle” chart. From the highest to the underside, it represents the decision hierarchy. 

On the very high: Python is executing our script ().

Subsequent: the script’s __main__ execution (:1()). Then the operate run_all_systems. Inside that, it calls two key features: iteration_heavy_task and cpu_heavy_task.

The memory-intensive processing half isn’t labelled on the chart. That’s as a result of the proportion of time related to this job is way smaller than the instances apportioned to the opposite two intensive features. In consequence, we see a a lot smaller, unlabelled block to the correct of the cpu_heavy_task block.

Observe that, for evaluation, there’s additionally a Snakeviz chart type referred to as a Sunburst chart. It seems to be a bit like a pie chart besides it comprises a set of more and more giant concentric circles and arcs. The thought beng that the time taken by features to run is represented by the angular extent of the arc measurement of the circle. The foundation operate is a circle in the course of viz. The foundation operate runs by calling the sub-functions beneath it and so forth. We wont be that show sort on this article.

Visible affirmation, like this, might be a lot extra impactful than looking at a desk of numbers. I didn’t have to guess anymore the place to look; the information was staring me proper within the face. 

The visualisation is shortly adopted by a block of textual content detailing the timings for numerous elements of your code, very like the output of the cprofile software. I’m solely displaying the primary dozen or so traces of this, as there have been 30+ in complete.

ncalls tottime percall cumtime percall filename:lineno(operate)
----------------------------------------------------------------
1 9.581 9.581 14.3 14.3 1062495604.py:34(iteration_heavy_task)
1 7.868 7.868 12.92 12.92 1062495604.py:14(cpu_heavy_task)
171796964 4.717 2.745e-08 4.717 2.745e-08 1062495604.py:31(simulate_tiny_op)
1 3.848 3.848 3.848 3.848 1062495604.py:22(memory_heavy_string_task)
34552942 1.91 5.527e-08 1.91 5.527e-08 ~:0()
34552942 1.836 5.313e-08 1.836 5.313e-08 ~:0()
34552942 1.305 3.778e-08 1.305 3.778e-08 ~:0()
1 0.02127 0.02127 31.09 31.09 :1()
4 0.0001764 4.409e-05 0.0001764 4.409e-05 socket.py:626(ship)
10 0.000123 1.23e-05 0.0004568 4.568e-05 iostream.py:655(write)
4 4.594e-05 1.148e-05 0.0002735 6.838e-05 iostream.py:259(schedule)
...
...
...

Step 3: The Repair

In fact, instruments like cprofiler and snakeviz don’t let you know how to kind out your efficiency points, however now that I knew precisely the place the issues had been, I may apply focused fixes. 

# final_showcase_fixed_v2.py
import time
import math
import numpy as np

# ===================================================================
CPU_ITERATIONS = 34552942
STRING_ITERATIONS = 46658100
LOOP_ITERATIONS = 171796964
# ===================================================================

# --- Repair 1: Vectorization for the CPU-Sure Process ---
def cpu_heavy_task_fixed(iterations):
    """
    Fastened through the use of NumPy to carry out the advanced math on a complete array
    directly, in extremely optimized C code as a substitute of a Python loop.
    """
    print("  -> Operating CPU-bound job...")
    # Create an array of numbers from 0 to iterations-1
    i = np.arange(iterations, dtype=np.float64)
    # The identical calculation, however vectorized, is orders of magnitude sooner
    result_array = np.sin(i) * np.cos(i) + np.sqrt(i)
    return np.sum(result_array)

# --- Repair 2: Environment friendly String Becoming a member of ---
def memory_heavy_string_task_fixed(iterations):
    """
    Fastened through the use of a listing comprehension and a single, environment friendly ''.be part of() name.
    This avoids creating hundreds of thousands of intermediate string objects.
    """
    print("  -> Operating Reminiscence/String-bound job...")
    chunk = "report_item_abcdefg_123456789_"
    # A listing comprehension is quick and memory-efficient
    elements = [f"|{chunk}{i}" for i in range(iterations)]
    return "".be part of(elements)

# --- Repair 3: Eliminating the "Thousand Cuts" Loop ---
def iteration_heavy_task_fixed(iterations):
    """
    Fastened by recognizing the duty generally is a no-op or a bulk operation.
    In a real-world state of affairs, you'll discover a solution to keep away from the loop solely.
    Right here, we reveal the repair by merely eradicating the pointless loop.
    The purpose is to point out the price of the loop itself was the issue.
    """
    print("  -> Operating Iteration-bound job...")
    # The repair is to discover a bulk operation or eradicate the necessity for the loop.
    # Because the authentic operate did nothing, the repair is to do nothing, however sooner.
    return "OK"

# --- Predominant Orchestrator ---
def run_all_systems():
    """
    The primary orchestrator now calls the FAST variations of the duties.
    """
    print("--- Beginning FINAL FAST Balanced Showcase ---")
    
    cpu_result = cpu_heavy_task_fixed(iterations=CPU_ITERATIONS)
    string_result = memory_heavy_string_task_fixed(iterations=STRING_ITERATIONS)
    iteration_result = iteration_heavy_task_fixed(iterations=LOOP_ITERATIONS)

    print("--- FINAL FAST Balanced Showcase Completed ---")

Now we are able to rerun the cprofiler on our up to date code.

import cProfile, pstats, io

pr = cProfile.Profile()
pr.allow()

# Run the operate you need to profile
run_all_systems()

pr.disable()

# Dump stats to a string and print the highest 10 by cumulative time
s = io.StringIO()
ps = pstats.Stats(pr, stream=s).sort_stats("cumtime")
ps.print_stats(10)
print(s.getvalue())

#
# begin of output
#

--- Beginning FINAL FAST Balanced Showcase ---
  -> Operating CPU-bound job...
  -> Operating Reminiscence/String-bound job...
  -> Operating Iteration-bound job...
--- FINAL FAST Balanced Showcase Completed ---
         197 operate calls in 6.063 seconds

   Ordered by: cumulative time
   Listing decreased from 52 to 10 resulting from restriction <10>

   ncalls  tottime  percall  cumtime  percall filename:lineno(operate)
        2    0.000    0.000    6.063    3.031 /dwelling/tom/.native/lib/python3.10/site-packages/IPython/core/interactiveshell.py:3541(run_code)
        2    0.000    0.000    6.063    3.031 {built-in methodology builtins.exec}
        1    0.002    0.002    6.063    6.063 /tmp/ipykernel_173802/1803406806.py:1()
        1    0.402    0.402    6.061    6.061 /tmp/ipykernel_173802/3782967348.py:52(run_all_systems)
        1    0.000    0.000    5.152    5.152 /tmp/ipykernel_173802/3782967348.py:27(memory_heavy_string_task_fixed)
        1    4.135    4.135    4.135    4.135 /tmp/ipykernel_173802/3782967348.py:35()
        1    1.017    1.017    1.017    1.017 {methodology 'be part of' of 'str' objects}
        1    0.446    0.446    0.505    0.505 /tmp/ipykernel_173802/3782967348.py:14(cpu_heavy_task_fixed)
        1    0.045    0.045    0.045    0.045 {built-in methodology numpy.arange}
        1    0.000    0.000    0.014    0.014 <__array_function__ internals>:177(sum)

That’s a improbable outcome that demonstrates the facility of profiling. We spent our effort on the elements of the code that mattered. To be thorough, I additionally ran snakeviz on the fastened script.

%%snakeviz
run_all_systems()

Essentially the most notable change is the discount in complete runtime, from roughly 30 seconds to roughly 6 seconds. This can be a 5x speedup, achieved by addressing the three important bottlenecks that had been seen within the “earlier than” profile.

Let’s have a look at every one individually.

1. The iteration_heavy_task

Earlier than (The Drawback)
Within the first picture, the big bar on the left, iteration_heavy_task, is the only greatest bottleneck, consuming 14.3 seconds.

• Why was it sluggish? This job was a traditional “dying by a thousand cuts.” The operate simulate_tiny_op did virtually nothing, nevertheless it was referred to as hundreds of thousands of instances from inside a pure Python for loop. The immense overhead of the Python interpreter beginning and stopping a operate name repeatedly was all the supply of the slowness.

The Repair
The fastened model, iteration_heavy_task_fixed, recognised that the purpose could possibly be achieved with out the loop. In our showcase, this meant eradicating the pointless loop solely. In a real-world utility, this is able to contain discovering a single “bulk” operation to switch the iterative one.

After (The Outcome)
Within the second picture, the iteration_heavy_task bar is fully gone. It’s now so quick that its runtime is a tiny fraction of a second and is invisible on the chart. We efficiently eradicated a 14.3-second downside.

2. The cpu_heavy_task

Earlier than (The Drawback)
The second main bottleneck, clearly seen as the big orange bar on the correct, is cpu_heavy_task, which took 12.9 seconds.

• Why was it sluggish? Just like the iteration job, this operate was additionally restricted by the velocity of the Python for loop. Whereas the mathematics operations inside had been quick, the interpreter needed to course of every of the hundreds of thousands of calculations individually, which is extremely inefficient for numerical duties.

The Repair
The repair was vectorisation utilizing the NumPy library. As a substitute of utilizing a Python loop, cpu_heavy_task_fixed created a NumPy array and carried out all of the mathematical operations (np.sqrt, np.sin, and so on.) on all the array concurrently. These operations are executed in extremely optimised, pre-compiled C code, fully bypassing the sluggish Python interpreter loop.

After (The Outcome).
Identical to the primary bottleneck, the cpu_heavy_task bar has vanished from the “after” diagram. Its runtime was decreased from 12.9 seconds to some milliseconds.

3. The memory_heavy_string_task

Earlier than (The Drawback):
Within the first diagram, the memory-heavy_string_task was working, however its runtime was small in comparison with the opposite two bigger points, so it was relegated to the small, unlabeled sliver of house on the far proper. It was a comparatively minor difficulty.

The Repair
The repair for this job was to switch the inefficient report += “…” string concatenation with a way more environment friendly methodology: constructing a listing of all of the string elements after which calling “”.be part of() a single time on the finish.

After (The Outcome)
Within the second diagram, we see the results of our success. Having eradicated the 2 10+ second bottlenecks, the memory-heavy-string-task-fixed is now the new dominant bottleneck, accounting for 4.34 seconds of the whole 5.22-second runtime.

Snakeviz even lets us look inside this fastened operate. The brand new most vital contributor is the orange bar labelled (record comprehension), which takes 3.52 seconds. This means that even within the fastened code, probably the most time-consuming half is now the method of making the in depth record of strings in reminiscence earlier than they are often joined.

Abstract

This text offers a hands-on information to figuring out and resolving efficiency points in Python code, arguing that builders ought to utilise profiling instruments to measure efficiency as a substitute of counting on instinct or guesswork to pinpoint the supply of slowdowns.

I demonstrated a methodical workflow utilizing two key instruments:-

• cProfile: Python’s built-in profiler, used to assemble detailed knowledge on operate calls and execution instances.
• snakeviz: A visualisation software that turns cProfile’s knowledge into an interactive “icicle” chart, making it straightforward to visually determine which elements of the code are consuming probably the most time.

The article makes use of a case research of a intentionally sluggish script engineered with three distinct and important bottlenecks:

1. An iteration-bound job: A operate referred to as hundreds of thousands of instances in a loop, showcasing the efficiency value of Python’s operate name overhead (“dying by a thousand cuts”).
2. A CPU-bound job: A for loop performing hundreds of thousands of math calculations, highlighting the inefficiency of pure Python for heavy numerical work.
3. A memory-bound job: A big string constructed inefficiently utilizing repeated += concatenation.

By analysing the snakeviz output, I pinpointed these three issues and utilized focused fixes.

• The iteration bottleneck was fastened by eliminating the pointless loop.
• The CPU bottleneck was resolved with vectorisation utilizing NumPy, which executes mathematical operations in quick, compiled C code.
• The reminiscence bottleneck was fastened by appending string elements to a listing and utilizing a single, environment friendly “”.be part of() name.

These fixes resulted in a dramatic speedup, lowering the script’s runtime from over 30 seconds to only over 6 seconds. I concluded by demonstrating that, even after main points are resolved, the profiler can be utilized once more to determine new, smaller bottlenecks, illustrating that efficiency tuning is an iterative course of guided by measurement.