1# Inferno 2 3 4 5[TOC] 6 7## Description 8 9Inferno is a flamegraph generator for native (C/C++) Android apps. It was 10originally written to profile and improve surfaceflinger performance 11(Android compositor) but it can be used for any native Android application 12. You can see a sample report generated with Inferno 13[here](./report.html). Report are self-contained in HTML so they can be 14exchanged easily. 15 16Notice there is no concept of time in a flame graph since all callstack are 17merged together. As a result, the width of a flamegraph represents 100% of 18the number of samples and the height is related to the number of functions on 19the stack when sampling occurred. 20 21 22 23 24In the flamegraph featured above you can see the main thread of SurfaceFlinger. 25It is immediatly apparent that most of the CPU time is spent processing messages 26`android::SurfaceFlinger::onMessageReceived`. The most expensive task is to ask 27 the screen to be refreshed as `android::DisplayDevice::prepare` shows in orange 28. This graphic division helps to see what part of the program is costly and 29where a developer's effort to improve performances should go. 30 31## Example of bottleneck 32 33A flamegraph give you instant vision on the CPU cycles cost centers but 34it can also be used to find specific offenders. To find them, look for 35plateaus. It is easier to see an example: 36 37 38 39In the previous flamegraph, two 40plateaus (due to `android::BufferQueueCore::validateConsistencyLocked`) 41are immediately apparent. 42 43## How it works 44 45Inferno relies on simpleperf to record the callstack of a native application 46thousands of times per second. Simpleperf takes care of unwinding the stack 47either using frame pointer (recommended) or dwarf. At the end of the recording 48`simpleperf` also symbolize all IPs automatically. The record are aggregated and 49dumps dumped to a file `perf.data`. This file is pulled from the Android device 50and processed on the host by Inferno. The callstacks are merged together to 51visualize in which part of an app the CPU cycles are spent. 52 53## How to use it 54 55Open a terminal and from `simpleperf/scripts` directory type: 56``` 57./inferno.sh (on Linux/Mac) 58inferno.bat (on Windows) 59``` 60 61Inferno will collect data, process them and automatically open your web browser 62to display the HTML report. 63 64## Parameters 65 66You can select how long to sample for, the color of the node and many other 67things. Use `-h` to get a list of all supported parameters. 68 69``` 70./inferno.sh -h 71``` 72 73## Troubleshooting 74 75### Messy flame graph 76 77A healthy flame graph features a single call site at its base (see [here](./report.html)). 78If you don't see a unique call site like `_start` or `_start_thread` at the base 79from which all flames originate, something went wrong. : Stack unwinding may 80fail to reach the root callsite. These incomplete 81callstack are impossible to merge properly. By default Inferno asks 82 `simpleperf` to unwind the stack via the kernel and frame pointers. Try to 83 perform unwinding with dwarf `-du`, you can further tune this setting. 84 85 86### No flames 87 88If you see no flames at all or a mess of 1 level flame without a common base, 89this may be because you compiled without frame pointers. Make sure there is no 90` -fomit-frame-pointer` in your build config. Alternatively, ask simpleperf to 91collect data with dward unwinding `-du`. 92 93 94 95### High percentage of lost samples 96 97If simpleperf reports a lot of lost sample it is probably because you are 98unwinding with `dwarf`. Dwarf unwinding involves copying the stack before it is 99processed. Try to use frame pointer unwinding which can be done by the kernel 100and it much faster. 101 102The cost of frame pointer is negligible on arm64 parameter but considerable 103 on arm 32-bit arch (due to register pressure). Use a 64-bit build for better 104 profiling. 105 106### run-as: package not debuggable 107 108If you cannot run as root, make sure the app is debuggable otherwise simpleperf 109will not be able to profile it. 110
