basic-software-profiling-cpu-usage.md (18823B)
1 # Basic Software Profiling: CPU Usage 2 3 ## Date: 2026/10/17 4 5 Software today is much slower than it should be. I believe this is, in part, because most software developers do not understand performance. 6 7 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. 8 9 ## Tools 10 11 [Perf](https://en.wikipedia.org/wiki/Perf_%28Linux%29) 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. 12 13 [Flamegraph](https://github.com/brendangregg/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](https://git.laack.co/flamegraph) which doesn't add j\*\*\*s\*\*\*\*\* to generated SVGs, but apart from that they function identically. 14 15 ## Reading a Flamegraph 16 17 Look at the flamegraph and try to understand the performance characteristics of the program: 18 19 <object data="/images/ex1-fg.svg" type="image/svg+xml" width="100%"></object> 20 21 <details> 22 <summary>Source code for the program that generated this flamegraph</summary> 23 24 ```cpp 25 uint32_t square(uint32_t input) { 26 return input * input; 27 } 28 29 // helper function to show stack depths 30 uint32_t squareHelper(uint32_t input) { 31 uint32_t x = 100 * 100; 32 return square(input); 33 } 34 35 uint32_t factorial(uint32_t input) { 36 uint32_t acc = 1; 37 38 while(input > 1) { 39 acc *= input; 40 input -= 1; 41 } 42 return acc; 43 } 44 45 int main() { 46 47 uint32_t ub = std::sqrt(UINT32_MAX); 48 49 for(uint32_t i = 0; i < ub; ++i) { 50 for(uint32_t x = 0; x < 100000; ++x) { 51 square(i); 52 } 53 54 for(uint32_t x = 0; x < 100000; ++x) { 55 squareHelper(i); 56 } 57 58 for(uint32_t x = 0; x < 20000; ++x) { 59 factorial(x % 10); 60 } 61 } 62 } 63 ``` 64 65 </details> 66 67 The amount of time spent with the `main` function on the call stack is represented by the width of its bar. 68 69 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. 70 71 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. 72 73 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. 74 75 ## Perf 76 77 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. 78 79 This is the command I ran to generate the data: 80 81 ```bash 82 perf record -g ./ex1 83 ``` 84 85 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. 86 87 To view aggregate information about the data in the `perf.data` file we can run: 88 89 ```bash 90 perf report -g 91 92 + 100.00% 0.00% ex1 ex1 [.] _start 93 + 100.00% 0.00% ex1 libc.so.6 [.] __libc_start_main 94 + 100.00% 0.00% ex1 libc.so.6 [.] call_init (inlined) 95 + 100.00% 0.00% ex1 libc.so.6 [.] 0x00007f6302a27781 96 + 99.79% 39.46% ex1 ex1 [.] main 97 + 34.97% 19.94% ex1 ex1 [.] squareHelper(unsigned int) 98 + 29.99% 19.90% ex1 ex1 [.] square(unsigned int) 99 + 21.61% 20.69% ex1 ex1 [.] factorial(unsigned int) 100 ... 101 ``` 102 103 If you are interested in information about which functions are calling each other, and the associated on-CPU times, you can run this command: 104 105 ```bash 106 perf report -g --stdio 107 108 ... 109 110 |--88.87%--main 111 | | 112 | |--19.84%--squareHelper(unsigned int) 113 | | | 114 | | --4.99%--square(unsigned int) 115 | | 116 | |--19.55%--factorial(unsigned int) 117 | | 118 | --10.07%--square(unsigned int) 119 | 120 |--5.02%--squareHelper(unsigned int) 121 | 122 |--4.98%--square(unsigned int) 123 | 124 --1.13%--factorial(unsigned int) 125 ... 126 ``` 127 128 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. 129 130 The final command worth mentioning is: 131 132 ``` 133 perf stat ./ex1 134 135 Performance counter stats for './ex1': 136 137 0 context-switches:u # 0.0 cs/sec cs_per_second 138 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 139 132 page-faults:u # 4.1 faults/sec page_faults_per_second 140 32,436.48 msec task-clock:u # nan CPUs CPUs_utilized 141 356,775 branch-misses:u # 0.0 % branch_miss_rate (88.89%) 142 63,706,083,988 branches:u # 1964.0 M/sec branch_frequency (88.89%) 143 134,126,037,336 cpu-cycles:u # 4.1 GHz cycles_frequency (88.89%) 144 300,951,617,557 instructions:u # 2.2 instructions insn_per_cycle (88.89%) 145 TopdownL1 # 4.4 % tma_backend_bound 146 # 61.4 % tma_bad_speculation (88.89%) 147 # 0.0 % tma_frontend_bound (77.78%) 148 # 34.2 % tma_retiring (88.89%) 149 150 32.440411262 seconds time elapsed 151 152 32.234190000 seconds user 153 0.002939000 seconds sys 154 ``` 155 156 Which provides high level information about the program's entire execution. 157 158 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. 159 160 ## Case Study 161 162 ### Amfora 163 164 [Amfora](https://github.com/makew0rld/amfora) is a fancy terminal-based Gemini protocol browser with syntax highlighting. Gemini is a text-based protocol which primarily distributes [gemtext](https://geminiprotocol.net/docs/gemtext.gmi) 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. 165 166 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](https://git.laack.co/foreseen/), a Gemini protocol browser I developed, as the oracle because it fulfils both criteria. 167 168 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. 169 170 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: 171 172 ```python 173 with open("lf.txt", "w") as f: 174 for i in range(0,200000): 175 f.write(str(i) + "\n") 176 ``` 177 178 The resulting file was ~1.3MB in size. I then ran both programs killing each program with Ctrl-C at first draw. 179 180 ``` 181 perf stat amfora lf.txt 182 183 Performance counter stats for 'amfora lf.txt': 184 185 0 context-switches:u # 0.0 cs/sec cs_per_second 186 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 187 2,742 page-faults:u # 9306.7 faults/sec page_faults_per_second 188 294.63 msec task-clock:u # nan CPUs CPUs_utilized 189 1,052,499 branch-misses:u # 0.2 % branch_miss_rate (90.02%) 190 575,712,764 branches:u # 1954.1 M/sec branch_frequency (89.14%) 191 955,922,330 cpu-cycles:u # 3.2 GHz cycles_frequency (87.42%) 192 2,669,651,845 instructions:u # 2.8 instructions insn_per_cycle (88.36%) 193 TopdownL1 # 6.7 % tma_backend_bound 194 # 24.9 % tma_bad_speculation (90.11%) 195 # 16.4 % tma_frontend_bound (79.92%) 196 # 52.0 % tma_retiring (90.37%) 197 198 0.473990805 seconds time elapsed 199 200 0.272938000 seconds user 201 0.021266000 seconds sys 202 203 ``` 204 205 ``` 206 perf stat foreseen lf.txt 207 208 Performance counter stats for 'foreseen lf.txt': 209 210 0 context-switches:u # 0.0 cs/sec cs_per_second 211 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 212 8,936 page-faults:u # 104842.2 faults/sec page_faults_per_second 213 85.23 msec task-clock:u # nan CPUs CPUs_utilized 214 209,148 branch-misses:u # 0.1 % branch_miss_rate (88.30%) 215 176,970,330 branches:u # 2076.3 M/sec branch_frequency (88.29%) 216 248,919,966 cpu-cycles:u # 2.9 GHz cycles_frequency (88.29%) 217 736,194,375 instructions:u # 3.0 instructions insn_per_cycle (88.99%) 218 TopdownL1 # 3.9 % tma_backend_bound 219 # 29.7 % tma_bad_speculation (89.50%) 220 # 21.2 % tma_frontend_bound (77.84%) 221 # 45.2 % tma_retiring (88.33%) 222 223 0.354029324 seconds time elapsed 224 225 0.066861000 seconds user 226 0.021186000 seconds sys 227 ``` 228 229 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. 230 231 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: 232 233 ``` 234 perf stat amfora lf.gmi 235 236 Performance counter stats for 'amfora lf.gmi': 237 238 0 context-switches:u # 0.0 cs/sec cs_per_second 239 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 240 900,174 page-faults:u # 12090.4 faults/sec page_faults_per_second 241 74,453.35 msec task-clock:u # nan CPUs CPUs_utilized 242 229,042,926 branch-misses:u # 0.4 % branch_miss_rate (88.94%) 243 62,713,218,442 branches:u # 842.3 M/sec branch_frequency (88.85%) 244 153,193,058,619 cpu-cycles:u # 2.1 GHz cycles_frequency (88.97%) 245 306,159,421,078 instructions:u # 2.0 instructions insn_per_cycle (88.90%) 246 TopdownL1 # 0.6 % tma_backend_bound 247 # 97.8 % tma_bad_speculation (88.83%) 248 # -0.1 % tma_frontend_bound (77.86%) 249 # 1.7 % tma_retiring (88.91%) 250 251 28.983517177 seconds time elapsed 252 253 63.894262000 seconds user 254 8.207151000 seconds sys 255 ``` 256 257 ``` 258 perf stat foreseen lf.gmi 259 260 Performance counter stats for 'foreseen lf.gmi': 261 262 0 context-switches:u # 0.0 cs/sec cs_per_second 263 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 264 6,551 page-faults:u # 90396.7 faults/sec page_faults_per_second 265 72.47 msec task-clock:u # nan CPUs CPUs_utilized 266 154,134 branch-misses:u # 0.1 % branch_miss_rate (88.98%) 267 166,258,529 branches:u # 2294.2 M/sec branch_frequency (88.99%) 268 232,523,911 cpu-cycles:u # 3.2 GHz cycles_frequency (88.96%) 269 705,288,446 instructions:u # 3.0 instructions insn_per_cycle (88.93%) 270 TopdownL1 # 7.6 % tma_backend_bound 271 # 23.2 % tma_bad_speculation (88.96%) 272 # 18.8 % tma_frontend_bound (77.41%) 273 # 50.4 % tma_retiring (88.45%) 274 275 0.381682961 seconds time elapsed 276 277 0.058132000 seconds user 278 0.015949000 seconds sys 279 ``` 280 281 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. 282 283 <object data="/images/amfora-before.svg" type="image/svg+xml" width="100%"></object> 284 285 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. 286 287 Looking in the RenderGemini function for string concatenations we find: 288 289 ``` 290 rendered += fmt.Sprintf("[%s]", config.GetColorString("preformatted_text")) + 291 buf + fmt.Sprintf("[%s:%s:-]\r\n", config.GetColorString("regular_text"), config.GetColorString("bg")) 292 293 ... 294 295 rendered += buf + "\r\n" 296 297 ... 298 299 rendered += ren 300 301 ... 302 303 buf += strings.TrimSuffix(lines[i], "\r") + "\r\n" 304 ``` 305 306 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. 307 308 Fortunately, this is easy to solve by using a strings.Builder which doesn't have this issue, and that's [what I did](https://github.com/makew0rld/amfora/pull/370/changes). After this change, we can perform our profiling again to find: 309 310 ``` 311 perf stat amfora lf.gmi 312 313 Performance counter stats for 'amfora lf.gmi': 314 315 0 context-switches:u # 0.0 cs/sec cs_per_second 316 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 317 6,039 page-faults:u # 4871.9 faults/sec page_faults_per_second 318 1,239.55 msec task-clock:u # nan CPUs CPUs_utilized 319 3,631,996 branch-misses:u # 0.1 % branch_miss_rate (89.02%) 320 3,017,891,055 branches:u # 2434.7 M/sec branch_frequency (88.96%) 321 4,654,531,574 cpu-cycles:u # 3.8 GHz cycles_frequency (88.93%) 322 14,168,541,807 instructions:u # 3.1 instructions insn_per_cycle (88.81%) 323 TopdownL1 # 2.2 % tma_backend_bound 324 # 36.9 % tma_bad_speculation (88.80%) 325 # 18.5 % tma_frontend_bound (77.83%) 326 # 42.5 % tma_retiring (88.90%) 327 328 1.185772969 seconds time elapsed 329 330 1.193871000 seconds user 331 0.039062000 seconds sys 332 ``` 333 334 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. 335 336 <object data="/images/amfora-after.svg" type="image/svg+xml" width="100%"></object> 337 338 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. 339 340 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. 341 342 ## What's Next 343 344 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: 345 346 1. Off-CPU profiling 347 2. Memory profiling 348 3. Build time profiling