basic-software-profiling-cpu-usage.md (18796B)
1 # Basic Software Profiling: CPU Usage 2 3 Software today is much slower than it should be. I believe this is, in part, because most software developers do not understand performance. 4 5 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. 6 7 ## Tools 8 9 [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. 10 11 [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. 12 13 ## Reading a Flamegraph 14 15 Look at the flamegraph and try to understand the performance characteristics of the program: 16 17 <object data="/images/ex1-fg.svg" type="image/svg+xml" width="100%"></object> 18 19 <details> 20 <summary>Source code for the program that generated this flamegraph</summary> 21 22 ```cpp 23 uint32_t square(uint32_t input) { 24 return input * input; 25 } 26 27 // helper function to show stack depths 28 uint32_t squareHelper(uint32_t input) { 29 uint32_t x = 100 * 100; 30 return square(input); 31 } 32 33 uint32_t factorial(uint32_t input) { 34 uint32_t acc = 1; 35 36 while(input > 1) { 37 acc *= input; 38 input -= 1; 39 } 40 return acc; 41 } 42 43 int main() { 44 45 uint32_t ub = std::sqrt(UINT32_MAX); 46 47 for(uint32_t i = 0; i < ub; ++i) { 48 for(uint32_t x = 0; x < 100000; ++x) { 49 square(i); 50 } 51 52 for(uint32_t x = 0; x < 100000; ++x) { 53 squareHelper(i); 54 } 55 56 for(uint32_t x = 0; x < 20000; ++x) { 57 factorial(x % 10); 58 } 59 } 60 } 61 ``` 62 63 </details> 64 65 The amount of time spent with the `main` function on the call stack is represented by the width of its bar. 66 67 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. 68 69 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. 70 71 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. 72 73 ## Perf 74 75 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. 76 77 This is the command I ran to generate the data: 78 79 ```bash 80 perf record -g ./ex1 81 ``` 82 83 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. 84 85 To view aggregate information about the data in the `perf.data` file we can run: 86 87 ```bash 88 perf report -g 89 90 + 100.00% 0.00% ex1 ex1 [.] _start 91 + 100.00% 0.00% ex1 libc.so.6 [.] __libc_start_main 92 + 100.00% 0.00% ex1 libc.so.6 [.] call_init (inlined) 93 + 100.00% 0.00% ex1 libc.so.6 [.] 0x00007f6302a27781 94 + 99.79% 39.46% ex1 ex1 [.] main 95 + 34.97% 19.94% ex1 ex1 [.] squareHelper(unsigned int) 96 + 29.99% 19.90% ex1 ex1 [.] square(unsigned int) 97 + 21.61% 20.69% ex1 ex1 [.] factorial(unsigned int) 98 ... 99 ``` 100 101 If you are interested in information about which functions are calling each other, and the associated on-CPU times, you can run this command: 102 103 ```bash 104 perf report -g --stdio 105 106 ... 107 108 |--88.87%--main 109 | | 110 | |--19.84%--squareHelper(unsigned int) 111 | | | 112 | | --4.99%--square(unsigned int) 113 | | 114 | |--19.55%--factorial(unsigned int) 115 | | 116 | --10.07%--square(unsigned int) 117 | 118 |--5.02%--squareHelper(unsigned int) 119 | 120 |--4.98%--square(unsigned int) 121 | 122 --1.13%--factorial(unsigned int) 123 ... 124 ``` 125 126 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. 127 128 The final command worth mentioning is: 129 130 ``` 131 perf stat ./ex1 132 133 Performance counter stats for './ex1': 134 135 0 context-switches:u # 0.0 cs/sec cs_per_second 136 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 137 132 page-faults:u # 4.1 faults/sec page_faults_per_second 138 32,436.48 msec task-clock:u # nan CPUs CPUs_utilized 139 356,775 branch-misses:u # 0.0 % branch_miss_rate (88.89%) 140 63,706,083,988 branches:u # 1964.0 M/sec branch_frequency (88.89%) 141 134,126,037,336 cpu-cycles:u # 4.1 GHz cycles_frequency (88.89%) 142 300,951,617,557 instructions:u # 2.2 instructions insn_per_cycle (88.89%) 143 TopdownL1 # 4.4 % tma_backend_bound 144 # 61.4 % tma_bad_speculation (88.89%) 145 # 0.0 % tma_frontend_bound (77.78%) 146 # 34.2 % tma_retiring (88.89%) 147 148 32.440411262 seconds time elapsed 149 150 32.234190000 seconds user 151 0.002939000 seconds sys 152 ``` 153 154 Which provides high level information about the program's entire execution. 155 156 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. 157 158 ## Case Study 159 160 ### Amfora 161 162 [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. 163 164 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. 165 166 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. 167 168 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: 169 170 ```python 171 with open("lf.txt", "w") as f: 172 for i in range(0,200000): 173 f.write(str(i) + "\n") 174 ``` 175 176 The resulting file was ~1.3MB in size. I then ran both programs killing each program with Ctrl-C at first draw. 177 178 ``` 179 perf stat amfora lf.txt 180 181 Performance counter stats for 'amfora lf.txt': 182 183 0 context-switches:u # 0.0 cs/sec cs_per_second 184 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 185 2,742 page-faults:u # 9306.7 faults/sec page_faults_per_second 186 294.63 msec task-clock:u # nan CPUs CPUs_utilized 187 1,052,499 branch-misses:u # 0.2 % branch_miss_rate (90.02%) 188 575,712,764 branches:u # 1954.1 M/sec branch_frequency (89.14%) 189 955,922,330 cpu-cycles:u # 3.2 GHz cycles_frequency (87.42%) 190 2,669,651,845 instructions:u # 2.8 instructions insn_per_cycle (88.36%) 191 TopdownL1 # 6.7 % tma_backend_bound 192 # 24.9 % tma_bad_speculation (90.11%) 193 # 16.4 % tma_frontend_bound (79.92%) 194 # 52.0 % tma_retiring (90.37%) 195 196 0.473990805 seconds time elapsed 197 198 0.272938000 seconds user 199 0.021266000 seconds sys 200 201 ``` 202 203 ``` 204 perf stat foreseen lf.txt 205 206 Performance counter stats for 'foreseen lf.txt': 207 208 0 context-switches:u # 0.0 cs/sec cs_per_second 209 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 210 8,936 page-faults:u # 104842.2 faults/sec page_faults_per_second 211 85.23 msec task-clock:u # nan CPUs CPUs_utilized 212 209,148 branch-misses:u # 0.1 % branch_miss_rate (88.30%) 213 176,970,330 branches:u # 2076.3 M/sec branch_frequency (88.29%) 214 248,919,966 cpu-cycles:u # 2.9 GHz cycles_frequency (88.29%) 215 736,194,375 instructions:u # 3.0 instructions insn_per_cycle (88.99%) 216 TopdownL1 # 3.9 % tma_backend_bound 217 # 29.7 % tma_bad_speculation (89.50%) 218 # 21.2 % tma_frontend_bound (77.84%) 219 # 45.2 % tma_retiring (88.33%) 220 221 0.354029324 seconds time elapsed 222 223 0.066861000 seconds user 224 0.021186000 seconds sys 225 ``` 226 227 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. 228 229 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: 230 231 ``` 232 perf stat amfora lf.gmi 233 234 Performance counter stats for 'amfora lf.gmi': 235 236 0 context-switches:u # 0.0 cs/sec cs_per_second 237 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 238 900,174 page-faults:u # 12090.4 faults/sec page_faults_per_second 239 74,453.35 msec task-clock:u # nan CPUs CPUs_utilized 240 229,042,926 branch-misses:u # 0.4 % branch_miss_rate (88.94%) 241 62,713,218,442 branches:u # 842.3 M/sec branch_frequency (88.85%) 242 153,193,058,619 cpu-cycles:u # 2.1 GHz cycles_frequency (88.97%) 243 306,159,421,078 instructions:u # 2.0 instructions insn_per_cycle (88.90%) 244 TopdownL1 # 0.6 % tma_backend_bound 245 # 97.8 % tma_bad_speculation (88.83%) 246 # -0.1 % tma_frontend_bound (77.86%) 247 # 1.7 % tma_retiring (88.91%) 248 249 28.983517177 seconds time elapsed 250 251 63.894262000 seconds user 252 8.207151000 seconds sys 253 ``` 254 255 ``` 256 perf stat foreseen lf.gmi 257 258 Performance counter stats for 'foreseen lf.gmi': 259 260 0 context-switches:u # 0.0 cs/sec cs_per_second 261 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 262 6,551 page-faults:u # 90396.7 faults/sec page_faults_per_second 263 72.47 msec task-clock:u # nan CPUs CPUs_utilized 264 154,134 branch-misses:u # 0.1 % branch_miss_rate (88.98%) 265 166,258,529 branches:u # 2294.2 M/sec branch_frequency (88.99%) 266 232,523,911 cpu-cycles:u # 3.2 GHz cycles_frequency (88.96%) 267 705,288,446 instructions:u # 3.0 instructions insn_per_cycle (88.93%) 268 TopdownL1 # 7.6 % tma_backend_bound 269 # 23.2 % tma_bad_speculation (88.96%) 270 # 18.8 % tma_frontend_bound (77.41%) 271 # 50.4 % tma_retiring (88.45%) 272 273 0.381682961 seconds time elapsed 274 275 0.058132000 seconds user 276 0.015949000 seconds sys 277 ``` 278 279 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. 280 281 <object data="/images/amfora-before.svg" type="image/svg+xml" width="100%"></object> 282 283 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. 284 285 Looking in the RenderGemini function for string concatenations, we find: 286 287 ``` 288 rendered += fmt.Sprintf("[%s]", config.GetColorString("preformatted_text")) + 289 buf + fmt.Sprintf("[%s:%s:-]\r\n", config.GetColorString("regular_text"), config.GetColorString("bg")) 290 291 ... 292 293 rendered += buf + "\r\n" 294 295 ... 296 297 rendered += ren 298 299 ... 300 301 buf += strings.TrimSuffix(lines[i], "\r") + "\r\n" 302 ``` 303 304 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. 305 306 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: 307 308 ``` 309 perf stat amfora lf.gmi 310 311 Performance counter stats for 'amfora lf.gmi': 312 313 0 context-switches:u # 0.0 cs/sec cs_per_second 314 0 cpu-migrations:u # 0.0 migrations/sec migrations_per_second 315 6,039 page-faults:u # 4871.9 faults/sec page_faults_per_second 316 1,239.55 msec task-clock:u # nan CPUs CPUs_utilized 317 3,631,996 branch-misses:u # 0.1 % branch_miss_rate (89.02%) 318 3,017,891,055 branches:u # 2434.7 M/sec branch_frequency (88.96%) 319 4,654,531,574 cpu-cycles:u # 3.8 GHz cycles_frequency (88.93%) 320 14,168,541,807 instructions:u # 3.1 instructions insn_per_cycle (88.81%) 321 TopdownL1 # 2.2 % tma_backend_bound 322 # 36.9 % tma_bad_speculation (88.80%) 323 # 18.5 % tma_frontend_bound (77.83%) 324 # 42.5 % tma_retiring (88.90%) 325 326 1.185772969 seconds time elapsed 327 328 1.193871000 seconds user 329 0.039062000 seconds sys 330 ``` 331 332 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. 333 334 <object data="/images/amfora-after.svg" type="image/svg+xml" width="100%"></object> 335 336 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. 337 338 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. 339 340 ## What's Next 341 342 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: 343 344 1. Off-CPU profiling 345 2. Memory profiling 346 3. Build time profiling