Basic Software Profiling: CPU Usage

Date: 2026/10/17

Software today is much slower than it should be. I believe this is, in part, because most software developers do not understand performance.

An easily explainable aspect of performance is profiling, and in this post I’ll discuss the basics of if. I’ll be focusing on compiled programs, but the core ideas may be applied to interpreted programs too.

Tools

Perf is a suite of tools for program profiling on Linux. We will mainly use it to collect aggregate performance metrics and call stack information. Analogous tools exist for other Free operating systems, and many bespoke tools exist for interpreted programs.

Flamegraph is a suite of tools for generating flamegraphs, a type of graph designed for the visualization of performance data. I’ll be using my fork of flamegraph which doesn’t add j***s***** to generated SVGs, but apart from that they function identically.

Reading a Flamegraph

Look at the flamegraph and try to understand the performance characteristics of the program:

Source code for the program that generated this flamegraph
uint32_t square(uint32_t input) {
    return input * input;
}

// helper function to show stack depths
uint32_t squareHelper(uint32_t input) {
    uint32_t x = 100 * 100;
    return square(input);
}

uint32_t factorial(uint32_t input) {
    uint32_t acc = 1;

    while(input > 1) {
        acc *= input;
        input -= 1;
    }
    return acc;
}

int main() {

    uint32_t ub = std::sqrt(UINT32_MAX);

    for(uint32_t i = 0; i < ub; ++i) {
        for(uint32_t x = 0; x < 100000; ++x) {
            square(i);
        }

        for(uint32_t x = 0; x < 100000; ++x) {
            squareHelper(i);
        }

        for(uint32_t x = 0; x < 20000; ++x) {
            factorial(x % 10);
        }
    }
}

The amount of time spent with the main function on the call stack is represented by the width of its bar.

Notice how there are multiple bars that say square. This is because we execute square at different call stack depths, and the width of each bar is proportional to the amount of time we spent with square on the call stack at a given stack depth.

Given how the widths are calculated, we can deduce the top-down width of a bar is how much time the program spent executing a specific function. In the case of squareHelper and square, we see we spent more time executing squareHelper than square, evidenced by the fact that from a top-down view, squareHelper is wider.

The astute among you might’ve noticed square and squareHelper seem to have been executed a few times outside of the main loop. This is due to how perf calculates the stack depth. The specifics of this are more in the weeds than we’ll discuss here, but it shouldn’t meaningfully impact the high level information that can be obtained from a flamegraph.

Perf

I’ve been loose with my words in terms of the specific data being displayed by the flamegraph. The command I used to generate the data for the flamegraph doesn’t technically track the amount of time spent with a function on the stack at a given depth, we are instead tracking the number of on-CPU samples taken with the function on the top of the call stack. This distinction is important because things like blocking I/O will not be reflected in our timings.

This is the command I ran to generate the data:

perf record -g ./ex1

The perf record command is used to run a command and record its performance profile, which is written to the perf.data file. By default, perf record will sample at 1000Hz while the program is on-CPU. The final option, -g, is used to specify we are interested in having the call stack included in our data.

To view aggregate information about the data in the perf.data file we can run:

perf report -g

+  100.00%     0.00%  ex1      ex1                   [.] _start
+  100.00%     0.00%  ex1      libc.so.6             [.] __libc_start_main
+  100.00%     0.00%  ex1      libc.so.6             [.] call_init (inlined)
+  100.00%     0.00%  ex1      libc.so.6             [.] 0x00007f6302a27781
+   99.79%    39.46%  ex1      ex1                   [.] main
+   34.97%    19.94%  ex1      ex1                   [.] squareHelper(unsigned int)
+   29.99%    19.90%  ex1      ex1                   [.] square(unsigned int)
+   21.61%    20.69%  ex1      ex1                   [.] factorial(unsigned int)
...

If you are interested in information about which functions are calling each other, and the associated on-CPU times, you can run this command:

perf report -g --stdio

...

|--88.87%--main
|          |
|          |--19.84%--squareHelper(unsigned int)
|          |          |
|          |           --4.99%--square(unsigned int)
|          |
|          |--19.55%--factorial(unsigned int)
|          |
|           --10.07%--square(unsigned int)
|
|--5.02%--squareHelper(unsigned int)
|
|--4.98%--square(unsigned int)
|
--1.13%--factorial(unsigned int)
...

I often find the --stdio command difficult to read because the report can get wide for non-trivial programs, and horizontal navigation in the terminal is unergonomic.

The final command worth mentioning is:

perf stat ./ex1

 Performance counter stats for './ex1':

                 0      context-switches:u              #      0.0 cs/sec  cs_per_second
                 0      cpu-migrations:u                #      0.0 migrations/sec  migrations_per_second
               132      page-faults:u                   #      4.1 faults/sec  page_faults_per_second
         32,436.48 msec task-clock:u                    #      nan CPUs  CPUs_utilized
           356,775      branch-misses:u                 #      0.0 %  branch_miss_rate           (88.89%)
    63,706,083,988      branches:u                      #   1964.0 M/sec  branch_frequency       (88.89%)
   134,126,037,336      cpu-cycles:u                    #      4.1 GHz  cycles_frequency         (88.89%)
   300,951,617,557      instructions:u                  #      2.2 instructions  insn_per_cycle  (88.89%)
         TopdownL1                                      #      4.4 %  tma_backend_bound
                                                        #     61.4 %  tma_bad_speculation        (88.89%)
                                                        #      0.0 %  tma_frontend_bound         (77.78%)
                                                        #     34.2 %  tma_retiring               (88.89%)

      32.440411262 seconds time elapsed

      32.234190000 seconds user
       0.002939000 seconds sys

Which provides high level information about the program’s entire execution.

This is the command I run most frequently when validating optimizations that target functions making up a sizeable percentage of the program’s CPU usage. If the program is uniformly expensive across many functions, perf stat is less useful than a full report / flamegraph.

Case Study

Amfora

Amfora is a fancy terminal-based Gemini protocol browser with syntax highlighting. Gemini is a text-based protocol which primarily distributes gemtext documents. These documents use a similar syntax to markdown, but it’s slightly simpler. I chose to profile this program because software in the Gemini ecosystem is often simple to understand, making it great for educational use cases.

When I start profiling software I often find it useful to use an oracle to understand what’s possible. An oracle in this context is a program that performs comparable tasks with decent performance characteristics. I chose Foreseen, a Gemini protocol browser I developed, as the oracle because it fulfils both criteria.

I like to start by searching for worst case performance scenarios. Worst cases are often an amplification of an average case, so by improving worst case performance we can often improve the average case.

To start our quest to find worse case scenarios for Amfora, we’ll load a large text file from disk, killing the program once the first lines of text are drawn in the terminal. Since neither browser supports text streaming, they both have to load and parse the document and then send text to the terminal, making this is a fair comparison. To generate a baseline document I executed the following Python code:

with open("lf.txt", "w") as f:
    for i in range(0,200000):
        f.write(str(i) + "\n")

The resulting file was ~1.3MB in size. I then ran both programs killing each program with Ctrl-C at first draw.

perf stat amfora lf.txt

Performance counter stats for 'amfora lf.txt':

                 0      context-switches:u               #      0.0 cs/sec  cs_per_second
                 0      cpu-migrations:u                 #      0.0 migrations/sec  migrations_per_second
             2,742      page-faults:u                    #   9306.7 faults/sec  page_faults_per_second
            294.63 msec task-clock:u                     #      nan CPUs  CPUs_utilized
         1,052,499      branch-misses:u                  #      0.2 %  branch_miss_rate           (90.02%)
       575,712,764      branches:u                       #   1954.1 M/sec  branch_frequency       (89.14%)
       955,922,330      cpu-cycles:u                     #      3.2 GHz  cycles_frequency         (87.42%)
     2,669,651,845      instructions:u                   #      2.8 instructions  insn_per_cycle  (88.36%)
         TopdownL1                                       #      6.7 %  tma_backend_bound
                                                         #     24.9 %  tma_bad_speculation        (90.11%)
                                                         #     16.4 %  tma_frontend_bound         (79.92%)
                                                         #     52.0 %  tma_retiring               (90.37%)

       0.473990805 seconds time elapsed

       0.272938000 seconds user
       0.021266000 seconds sys
perf stat foreseen lf.txt

 Performance counter stats for 'foreseen lf.txt':

                 0      context-switches:u               #      0.0 cs/sec  cs_per_second
                 0      cpu-migrations:u                 #      0.0 migrations/sec  migrations_per_second
             8,936      page-faults:u                    # 104842.2 faults/sec  page_faults_per_second
             85.23 msec task-clock:u                     #      nan CPUs  CPUs_utilized
           209,148      branch-misses:u                  #      0.1 %  branch_miss_rate           (88.30%)
       176,970,330      branches:u                       #   2076.3 M/sec  branch_frequency       (88.29%)
       248,919,966      cpu-cycles:u                     #      2.9 GHz  cycles_frequency         (88.29%)
       736,194,375      instructions:u                   #      3.0 instructions  insn_per_cycle  (88.99%)
         TopdownL1                                       #      3.9 %  tma_backend_bound
                                                         #     29.7 %  tma_bad_speculation        (89.50%)
                                                         #     21.2 %  tma_frontend_bound         (77.84%)
                                                         #     45.2 %  tma_retiring               (88.33%)

       0.354029324 seconds time elapsed

       0.066861000 seconds user
       0.021186000 seconds sys

This is very unscientific, especially given I was watching to see when to kill the programs, but this is how I normally start profiling. We can’t conclude anything from these results as the timings are within margin of error.

Since Amfora can highlight code let’s see how that impacts performance. I pasted the Python code above at the top of the text file, encapsulated by gemtext preformatted text lines, and then renamed the file to lf.gmi. We then find:

perf stat amfora lf.gmi

 Performance counter stats for 'amfora lf.gmi':

                 0      context-switches:u               #      0.0 cs/sec  cs_per_second
                 0      cpu-migrations:u                 #      0.0 migrations/sec  migrations_per_second
           900,174      page-faults:u                    #  12090.4 faults/sec  page_faults_per_second
         74,453.35 msec task-clock:u                     #      nan CPUs  CPUs_utilized
       229,042,926      branch-misses:u                  #      0.4 %  branch_miss_rate           (88.94%)
    62,713,218,442      branches:u                       #    842.3 M/sec  branch_frequency       (88.85%)
   153,193,058,619      cpu-cycles:u                     #      2.1 GHz  cycles_frequency         (88.97%)
   306,159,421,078      instructions:u                   #      2.0 instructions  insn_per_cycle  (88.90%)
         TopdownL1                                       #      0.6 %  tma_backend_bound
                                                         #     97.8 %  tma_bad_speculation        (88.83%)
                                                         #     -0.1 %  tma_frontend_bound         (77.86%)
                                                         #      1.7 %  tma_retiring               (88.91%)

      28.983517177 seconds time elapsed

      63.894262000 seconds user
       8.207151000 seconds sys
perf stat foreseen lf.gmi

 Performance counter stats for 'foreseen lf.gmi':

                 0      context-switches:u               #      0.0 cs/sec  cs_per_second
                 0      cpu-migrations:u                 #      0.0 migrations/sec  migrations_per_second
             6,551      page-faults:u                    #  90396.7 faults/sec  page_faults_per_second
             72.47 msec task-clock:u                     #      nan CPUs  CPUs_utilized
           154,134      branch-misses:u                  #      0.1 %  branch_miss_rate           (88.98%)
       166,258,529      branches:u                       #   2294.2 M/sec  branch_frequency       (88.99%)
       232,523,911      cpu-cycles:u                     #      3.2 GHz  cycles_frequency         (88.96%)
       705,288,446      instructions:u                   #      3.0 instructions  insn_per_cycle  (88.93%)
         TopdownL1                                       #      7.6 %  tma_backend_bound
                                                         #     23.2 %  tma_bad_speculation        (88.96%)
                                                         #     18.8 %  tma_frontend_bound         (77.41%)
                                                         #     50.4 %  tma_retiring               (88.45%)

       0.381682961 seconds time elapsed

       0.058132000 seconds user
       0.015949000 seconds sys

That’s a lot slower for Amfora, ~76x slower than Foreseen, and ~61x slower than it was with the normal text file. Let’s look at a flamegraph for this.

So 68.34% of samples occurred while executing the GC. That’s a bit surprising as there shouldn’t be that many allocations occuring. We also see 25.47% of the samples occur when executing the RenderGemini function. Of this, most of it is spent on concatstring3, a function used to concatenate strings.

Looking in the RenderGemini function for string concatenations we find:

rendered += fmt.Sprintf("[%s]", config.GetColorString("preformatted_text")) +
    buf + fmt.Sprintf("[%s:%s:-]\r\n", config.GetColorString("regular_text"), config.GetColorString("bg"))

...

rendered += buf + "\r\n"

...

rendered += ren

...

buf += strings.TrimSuffix(lines[i], "\r") + "\r\n"

Yeah, this isn’t great. Each of these seemingly innocuous appends is very costly. Go strings are immutable so each append creates an entirely new string, and since we accumulate into buf until reaching a preformatted text switch line (“```” lines), we have to re-create the accumulator string 200,000 times to parse our file.

Fortunately, this is easy to solve by using a strings.Builder which doesn’t have this issue, and that’s what I did. After this change, we can perform our profiling again to find:

perf stat amfora lf.gmi

 Performance counter stats for 'amfora lf.gmi':

                 0      context-switches:u               #      0.0 cs/sec  cs_per_second
                 0      cpu-migrations:u                 #      0.0 migrations/sec  migrations_per_second
             6,039      page-faults:u                    #   4871.9 faults/sec  page_faults_per_second
          1,239.55 msec task-clock:u                     #      nan CPUs  CPUs_utilized
         3,631,996      branch-misses:u                  #      0.1 %  branch_miss_rate           (89.02%)
     3,017,891,055      branches:u                       #   2434.7 M/sec  branch_frequency       (88.96%)
     4,654,531,574      cpu-cycles:u                     #      3.8 GHz  cycles_frequency         (88.93%)
    14,168,541,807      instructions:u                   #      3.1 instructions  insn_per_cycle  (88.81%)
         TopdownL1                                       #      2.2 %  tma_backend_bound
                                                         #     36.9 %  tma_bad_speculation        (88.80%)
                                                         #     18.5 %  tma_frontend_bound         (77.83%)
                                                         #     42.5 %  tma_retiring               (88.90%)

       1.185772969 seconds time elapsed

       1.193871000 seconds user
       0.039062000 seconds sys

This is much better. We are now using ~33x less cycles, and it’s ~25x faster than before. It’s also only 3x slower than foreseen on this task.

Looking again at the flamegraph, we see the RenderGemini function is now mostly performing word wrapping, and the GC is only utilizing the CPU during 16.48% of our samples.

We could continue identifying hotspots, but now there are 4 ~equally costly groupings, and as performance becomes more diffuse, optimizations will generally have a smaller impact on overall performance, so I’m going to leave this as is.

What’s Next

This post only discussed on-CPU profiling for program execution. There are a few additional metrics worth profiling that I’ll discuss in the future:

  1. Off-CPU profiling
  2. Memory profiling
  3. Build time profiling