1# Profcollect 2 3Profcollect is a system daemon that facilitates sampling profile collection and reporting for native 4platform applications. 5 6Profcollect can only be enabled on `userdebug` or `eng` builds. 7 8## Supported Platforms 9 10Currently Profcollect only supports collecting profiles from Coresight ETM enabled ARM devices. 11 12Instructions to enable Coresight ETM can be found from the 13[simpleperf manual](https://android.googlesource.com/platform/system/extras/+/refs/heads/master/simpleperf/doc/collect_etm_data_for_autofdo.md). 14 15## Usage 16 17Profcollect has two components: `profcollectd`, the system daemon, and `profcollectctl`, the command 18line interface. 19 20### Collection 21 22`profcollectd` can be started from `adb` directly (under root), or automatically on system boot by 23setting system property through: 24 25``` 26adb shell device_config put profcollect_native_boot enabled true 27``` 28 29Profcollect collects profiles periodically, as well as through triggers like app launch events. Only 30a percentage of these events result in a profile collection to avoid using too much resource, these 31are controlled by the following configurations: 32 33| Event | Config | 34|------------|------------------------| 35| Periodic | collection\_interval | 36| App launch | applaunch\_trace\_freq | 37 38Setting the frequency value to `0` disables collection for the corresponding event. 39 40#### Custom configuration 41 42In adb root: 43 44``` 45# Record every 60s (By default, record every 10m). The actual interval will be longer than the 46# set value if the device goes to hibernation. 47oriole:/ # setprop persist.device_config.profcollect_native_boot.collection_interval 60 48 49# Each time recording, record ETM data for 1s (By default, it's 0.5s). 50oriole:/ # setprop persist.device_config.profcollect_native_boot.sampling_period 1000 51 52# Set ETM data storage limit to 50G (By default, it is 512M). 53oriole:/ # setprop persist.device_config.profcollect_native_boot.max_trace_limit 53687091200 54 55# Enable ETM data collection (By default, it's decided by the server). 56oriole:/ # setprop persist.device_config.profcollect_native_boot.enabled true 57 58# After adjusting configuration, need to restart profcollectd 59oriole:/ # setprop ctl.stop profcollectd 60# Wait for a few seconds. 61oriole:/ # setprop ctl.start profcollectd 62 63# Check if profcollectd is running 64oriole:/ # ps -e | grep profcollectd 65root 918 1 10945660 47040 binder_wait_for_work 0 S profcollectd 66 67# Check if the new configuration takes effect. 68oriole:/ # cat /data/misc/profcollectd/output/config.json 69{"version":1,"node_id":[189,15,145,225,97,167],"build_fingerprint":"google/oriole/oriole:Tiramisu/TP1A.220223.002/8211650:userdebug/dev-keys","collection_interval":{"secs":60,"nanos":0},"sampling_period":{"secs":1,"nanos":0},"binary_filter":"^/(system|apex/.+)/(bin|lib|lib64)/.+","max_trace_limit":53687091200} 70``` 71 72To check existing collected ETM data: 73``` 74oriole:/ # cd data/misc/profcollectd/trace/ 75oriole:/data/misc/profcollectd/trace # ls 76``` 77 78To check if ETM data can be collected successfully: 79``` 80# Trigger one collection manually. 81oriole:/ # profcollectctl once 82Trace once 83 84# Check trace directory to see if there is a recent manual trace file. 85oriole:/ # ls /data/misc/profcollectd/trace/ 8620220224-222946_manual.etmtrace 87``` 88 89If there are too many trace files, we need to processing them to avoid reaching storage limit. 90It may take a long time. 91``` 92oriole:/ # profcollectctl process 93Processing traces 94``` 95 96### Processing 97 98The raw tracing data needs to be combined with the original binary to create the AutoFDO branch 99list. This is a costly process, thus it is done separately from the profile collection. Profcollect 100attempts to process all the traces when the device is idle and connected to a power supply. It can 101also be initiated by running: 102 103``` 104adb shell profcollectctl process 105``` 106 107### Reporting 108 109#### Manual 110 111After actively using the device for a period of time, the device should have gathered enough data to 112generate a good quality PGO profile that represents typical system usage. Run the following command 113to create a profile report: 114 115``` 116$ adb shell profcollectctl report 117Creating profile report 118Report created at: 12345678-0000-abcd-8000-12345678abcd 119``` 120 121You can then fetch the report by running (under root): 122 123``` 124adb pull /data/misc/profcollectd/report/12345678-0000-abcd-8000-12345678abcd.zip 125``` 126 127#### Automated Uploading to Server 128 129*In development* 130 131### Post Processing 132 133For each trace file, run: 134 135``` 136simpleperf inject \ 137 -i {TRACE_FILE_NAME} \ 138 -o {OUTPUT_FILE_NAME}.data \ 139 --binary {BINARY_NAME} \ 140 --symdir out/target/product/{PRODUCT_NAME}/symbols 141``` 142 143Afterwards, run [AutoFDO](https://github.com/google/autofdo) to generate Clang PGO profiles: 144 145``` 146create_llvm_prof \ 147 --profiler text \ 148 --binary=${BINARY_PATH} \ 149 --profile=${INPUT_FILE_NAME} \ 150 --out={OUTPUT_FILE_NAME}.profdata 151``` 152 153Finally, merge all the PGO profiles into one profile: 154 155``` 156find {INPUT_DIR} -name *.profdata > proflist 157prebuilts/clang/host/linux-x86/llvm-binutils-stable/llvm-profdata merge \ 158 --binary \ 159 --sample \ 160 --input-files proflist \ 161 --output merged.profdata 162``` 163 164More profile data usually generates better quality profiles. You may combine data from multiple 165devices running the same build to improve profile quality, and/or reduce the performance impact for 166each device (by reducing collection frequency). 167