basic-software-profiling-cpu-usage.gmi (18676B)
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 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 => https://en.wikipedia.org/wiki/Perf_%28Linux%29 14 15 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. 16 17 => https://github.com/brendangregg/FlameGraph 18 => https://git.laack.co/flamegraph 19 20 21 ## Reading a Flamegraph 22 23 Look at the flamegraph and try to understand the performance characteristics of the program: 24 25 => https://blog.laack.co/images/ex1-fg.svg 26 27 Source code for the program that generated this flamegraph: 28 29 ```cpp 30 uint32_t square(uint32_t input) { 31 return input * input; 32 } 33 34 // helper function to show stack depths 35 uint32_t squareHelper(uint32_t input) { 36 uint32_t x = 100 * 100; 37 return square(input); 38 } 39 40 uint32_t factorial(uint32_t input) { 41 uint32_t acc = 1; 42 43 while(input > 1) { 44 acc *= input; 45 input -= 1; 46 } 47 return acc; 48 } 49 50 int main() { 51 52 uint32_t ub = std::sqrt(UINT32_MAX); 53 54 for(uint32_t i = 0; i < ub; ++i) { 55 for(uint32_t x = 0; x < 100000; ++x) { 56 square(i); 57 } 58 59 for(uint32_t x = 0; x < 100000; ++x) { 60 squareHelper(i); 61 } 62 63 for(uint32_t x = 0; x < 20000; ++x) { 64 factorial(x % 10); 65 } 66 } 67 } 68 ``` 69 70 The amount of time spent with the `main` function on the call stack is represented by the width of its bar. 71 72 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. 73 74 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. 75 76 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. 77 78 ## Perf 79 80 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. 81 82 This is the command I ran to generate the data: 83 84 ```bash 85 perf record -g ./ex1 86 ``` 87 88 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. 89 90 To view aggregate information about the data in the `perf.data` file we can run: 91 92 ```bash 93 perf report -g 94 95 + 100.00% 0.00% ex1 ex1 [.] _start 96 + 100.00% 0.00% ex1 libc.so.6 [.] __libc_start_main 97 + 100.00% 0.00% ex1 libc.so.6 [.] call_init (inlined) 98 + 100.00% 0.00% ex1 libc.so.6 [.] 0x00007f6302a27781 99 + 99.79% 39.46% ex1 ex1 [.] main 100 + 34.97% 19.94% ex1 ex1 [.] squareHelper(unsigned int) 101 + 29.99% 19.90% ex1 ex1 [.] square(unsigned int) 102 + 21.61% 20.69% ex1 ex1 [.] factorial(unsigned int) 103 ... 104 ``` 105 106 If you are interested in information about which functions are calling each other, and the associated on-CPU times, you can run this command: 107 108 ```bash 109 perf report -g --stdio 110 111 ... 112 113 |--88.87%--main 114 | | 115 | |--19.84%--squareHelper(unsigned int) 116 | | | 117 | | --4.99%--square(unsigned int) 118 | | 119 | |--19.55%--factorial(unsigned int) 120 | | 121 | --10.07%--square(unsigned int) 122 | 123 |--5.02%--squareHelper(unsigned int) 124 | 125 |--4.98%--square(unsigned int) 126 | 127 --1.13%--factorial(unsigned int) 128 ... 129 ``` 130 131 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. 132 133 The final command worth mentioning is: 134 135 ``` 136 perf stat ./ex1 137 138 Performance counter stats for './ex1': 139 140 0 context-switches:u # 0.0 cs/sec cs_per_second 141 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 142 132 page-faults:u # 4.1 faults/sec page_faults_per_second 143 32,436.48 msec task-clock:u # nan CPUs CPUs_utilized 144 356,775 branch-misses:u # 0.0 % branch_miss_rate (88.89%) 145 63,706,083,988 branches:u # 1964.0 M/sec branch_frequency (88.89%) 146 134,126,037,336 cpu-cycles:u # 4.1 GHz cycles_frequency (88.89%) 147 300,951,617,557 instructions:u # 2.2 instructions insn_per_cycle (88.89%) 148 TopdownL1 # 4.4 % tma_backend_bound 149 # 61.4 % tma_bad_speculation (88.89%) 150 # 0.0 % tma_frontend_bound (77.78%) 151 # 34.2 % tma_retiring (88.89%) 152 153 32.440411262 seconds time elapsed 154 155 32.234190000 seconds user 156 0.002939000 seconds sys 157 ``` 158 159 Which provides high level information about the program's entire execution. 160 161 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. 162 163 ## Case Study 164 165 ### Amfora 166 167 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. 168 169 => https://github.com/makew0rld/amfora 170 => https://geminiprotocol.net/docs/gemtext.gmi 171 172 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. 173 174 => https://git.laack.co/foreseen/ 175 176 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. 177 178 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: 179 180 ```python 181 with open("lf.txt", "w") as f: 182 for i in range(0,200000): 183 f.write(str(i) + "\n") 184 ``` 185 186 The resulting file was ~1.3MB in size. I then ran both programs killing each program with Ctrl-C at first draw. 187 188 ``` 189 perf stat amfora lf.txt 190 191 Performance counter stats for 'amfora lf.txt': 192 193 0 context-switches:u # 0.0 cs/sec cs_per_second 194 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 195 2,742 page-faults:u # 9306.7 faults/sec page_faults_per_second 196 294.63 msec task-clock:u # nan CPUs CPUs_utilized 197 1,052,499 branch-misses:u # 0.2 % branch_miss_rate (90.02%) 198 575,712,764 branches:u # 1954.1 M/sec branch_frequency (89.14%) 199 955,922,330 cpu-cycles:u # 3.2 GHz cycles_frequency (87.42%) 200 2,669,651,845 instructions:u # 2.8 instructions insn_per_cycle (88.36%) 201 TopdownL1 # 6.7 % tma_backend_bound 202 # 24.9 % tma_bad_speculation (90.11%) 203 # 16.4 % tma_frontend_bound (79.92%) 204 # 52.0 % tma_retiring (90.37%) 205 206 0.473990805 seconds time elapsed 207 208 0.272938000 seconds user 209 0.021266000 seconds sys 210 211 ``` 212 213 ``` 214 perf stat foreseen lf.txt 215 216 Performance counter stats for 'foreseen lf.txt': 217 218 0 context-switches:u # 0.0 cs/sec cs_per_second 219 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 220 8,936 page-faults:u # 104842.2 faults/sec page_faults_per_second 221 85.23 msec task-clock:u # nan CPUs CPUs_utilized 222 209,148 branch-misses:u # 0.1 % branch_miss_rate (88.30%) 223 176,970,330 branches:u # 2076.3 M/sec branch_frequency (88.29%) 224 248,919,966 cpu-cycles:u # 2.9 GHz cycles_frequency (88.29%) 225 736,194,375 instructions:u # 3.0 instructions insn_per_cycle (88.99%) 226 TopdownL1 # 3.9 % tma_backend_bound 227 # 29.7 % tma_bad_speculation (89.50%) 228 # 21.2 % tma_frontend_bound (77.84%) 229 # 45.2 % tma_retiring (88.33%) 230 231 0.354029324 seconds time elapsed 232 233 0.066861000 seconds user 234 0.021186000 seconds sys 235 ``` 236 237 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. 238 239 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: 240 241 ``` 242 perf stat amfora lf.gmi 243 244 Performance counter stats for 'amfora lf.gmi': 245 246 0 context-switches:u # 0.0 cs/sec cs_per_second 247 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 248 900,174 page-faults:u # 12090.4 faults/sec page_faults_per_second 249 74,453.35 msec task-clock:u # nan CPUs CPUs_utilized 250 229,042,926 branch-misses:u # 0.4 % branch_miss_rate (88.94%) 251 62,713,218,442 branches:u # 842.3 M/sec branch_frequency (88.85%) 252 153,193,058,619 cpu-cycles:u # 2.1 GHz cycles_frequency (88.97%) 253 306,159,421,078 instructions:u # 2.0 instructions insn_per_cycle (88.90%) 254 TopdownL1 # 0.6 % tma_backend_bound 255 # 97.8 % tma_bad_speculation (88.83%) 256 # -0.1 % tma_frontend_bound (77.86%) 257 # 1.7 % tma_retiring (88.91%) 258 259 28.983517177 seconds time elapsed 260 261 63.894262000 seconds user 262 8.207151000 seconds sys 263 ``` 264 265 ``` 266 perf stat foreseen lf.gmi 267 268 Performance counter stats for 'foreseen lf.gmi': 269 270 0 context-switches:u # 0.0 cs/sec cs_per_second 271 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 272 6,551 page-faults:u # 90396.7 faults/sec page_faults_per_second 273 72.47 msec task-clock:u # nan CPUs CPUs_utilized 274 154,134 branch-misses:u # 0.1 % branch_miss_rate (88.98%) 275 166,258,529 branches:u # 2294.2 M/sec branch_frequency (88.99%) 276 232,523,911 cpu-cycles:u # 3.2 GHz cycles_frequency (88.96%) 277 705,288,446 instructions:u # 3.0 instructions insn_per_cycle (88.93%) 278 TopdownL1 # 7.6 % tma_backend_bound 279 # 23.2 % tma_bad_speculation (88.96%) 280 # 18.8 % tma_frontend_bound (77.41%) 281 # 50.4 % tma_retiring (88.45%) 282 283 0.381682961 seconds time elapsed 284 285 0.058132000 seconds user 286 0.015949000 seconds sys 287 ``` 288 289 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. 290 291 => https://blog.laack.co/images/amfora-before.svg 292 293 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. 294 295 Looking in the RenderGemini function for string concatenations we find: 296 297 ``` 298 rendered += fmt.Sprintf("[%s]", config.GetColorString("preformatted_text")) + 299 buf + fmt.Sprintf("[%s:%s:-]\r\n", config.GetColorString("regular_text"), config.GetColorString("bg")) 300 301 ... 302 303 rendered += buf + "\r\n" 304 305 ... 306 307 rendered += ren 308 309 ... 310 311 buf += strings.TrimSuffix(lines[i], "\r") + "\r\n" 312 ``` 313 314 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. 315 316 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: 317 318 => https://github.com/makew0rld/amfora/pull/370/changes 319 320 ``` 321 perf stat amfora lf.gmi 322 323 Performance counter stats for 'amfora lf.gmi': 324 325 0 context-switches:u # 0.0 cs/sec cs_per_second 326 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 327 6,039 page-faults:u # 4871.9 faults/sec page_faults_per_second 328 1,239.55 msec task-clock:u # nan CPUs CPUs_utilized 329 3,631,996 branch-misses:u # 0.1 % branch_miss_rate (89.02%) 330 3,017,891,055 branches:u # 2434.7 M/sec branch_frequency (88.96%) 331 4,654,531,574 cpu-cycles:u # 3.8 GHz cycles_frequency (88.93%) 332 14,168,541,807 instructions:u # 3.1 instructions insn_per_cycle (88.81%) 333 TopdownL1 # 2.2 % tma_backend_bound 334 # 36.9 % tma_bad_speculation (88.80%) 335 # 18.5 % tma_frontend_bound (77.83%) 336 # 42.5 % tma_retiring (88.90%) 337 338 1.185772969 seconds time elapsed 339 340 1.193871000 seconds user 341 0.039062000 seconds sys 342 ``` 343 344 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. 345 346 => https://blog.laack.co/images/amfora-after.svg 347 348 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. 349 350 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. 351 352 ## What's Next 353 354 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: 355 356 1. Off-CPU profiling 357 2. Memory profiling 358 3. Build time profiling