| Primiano Tucci | b808373 | 2018-04-09 15:38:37 +0100 | [diff] [blame] | 1 | # Perfetto benchmarks |
| 2 | |
| 3 | |
| 4 | *** note |
| 5 | **This doc is WIP**, stay tuned. |
| 6 | <!-- TODO(primiano): Summarize results of perfetto_benchmarks --> |
| 7 | *** |
| 8 | |
| 9 | This doc should show the charts of `perfetto_benchmarks`, showing cpu usage and |
| 10 | tracing bandwidth for both writing (producer->service) and reading |
| 11 | (service->file / service->consumer). |
| 12 | |
| 13 | In two modes: |
| 14 | - Measure peak tracing bandwidth saturating the cpu: the producer(s) write as |
| 15 | much data as they can, reaching 100% cpu usage. |
| 16 | - Measure CPU overhead vs constant bandwidth: the producer(s) writes data at a |
| 17 | pre-defined rate. |
| 18 | |
| 19 | Tweaking the various parameters, such as: |
| 20 | - Number of writers |
| 21 | - Size of the shared memory buffer |
| 22 | - Size of each TracePacket. |
| 23 | |
| 24 | **TL;DR:** |
| 25 | Peak producer-to-service tracing bandwidth: |
| 26 | * Linux desktop: ~1.3 GB/s |
| 27 | * Android Pixel: ~1 GB/s |
| 28 | |
| 29 | Producer-to-service CPU overhead when writing ~3 MB/s: 0.01 - 0.03 |
| 30 | (0.01 := 1% cpu time of one core) |
| 31 | |
| 32 | CPU overhead for translating ftrace raw pipe into protobuf: |
| 33 | * Android Pixel: 0.00-0.01 when idle. |
| 34 | * Android Pixel: 0.02-0.04 with 8 cores @ 8.0 CPU usage (raytracer). |
| 35 | * Linux desktop: TBD |
