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