Add performance.md and add instruction for linking tcmalloc

This commit is contained in:
Yilun Chong 2018-03-22 17:08:06 -07:00
parent 9dc0a4d5cf
commit 745ef89ebf
2 changed files with 318 additions and 1 deletions

View File

@ -3,7 +3,9 @@
This directory contains benchmarking schemas and data sets that you This directory contains benchmarking schemas and data sets that you
can use to test a variety of performance scenarios against your can use to test a variety of performance scenarios against your
protobuf language runtime. protobuf language runtime. If you are looking for performance
numbers of officially support languages, see [here](
https://github.com/google/protobuf/blob/master/docs/Performance.md)
## Prerequisite ## Prerequisite
@ -17,6 +19,11 @@ We are using [google/benchmark](https://github.com/google/benchmark) as the
benchmark tool for testing cpp. This will be automaticly made during build the benchmark tool for testing cpp. This will be automaticly made during build the
cpp benchmark. cpp benchmark.
The cpp protobuf performance can be improved by linking with [tcmalloc library](
https://gperftools.github.io/gperftools/tcmalloc.html). For using tcmalloc, you
need to build [gpertools](https://github.com/gperftools/gperftools) to generate
libtcmallc.so library.
### Java ### Java
We're using maven to build the java benchmarks, which is the same as to build We're using maven to build the java benchmarks, which is the same as to build
the Java protobuf. There're no other tools need to install. We're using the Java protobuf. There're no other tools need to install. We're using
@ -64,6 +71,12 @@ $ make java
$ make cpp $ make cpp
``` ```
For linking with tcmalloc:
```
$ env LD_PRELOAD={directory to libtcmalloc.so} make cpp
```
### Python: ### Python:
We have three versions of python protobuf implementation: pure python, cpp We have three versions of python protobuf implementation: pure python, cpp

304
docs/performance.md Normal file
View File

@ -0,0 +1,304 @@
# Protobuf Perforamcne
This benchmark result is tested on workstation with processor of Intel® Xeon® Processor E5-2630 and 32GB RAM
This table contains 3 languages' results:
* **C++** - For C++ there're 3 kinds of parsing ways:
* **new** - This is for using new operator for creating message instance.
* **new arena** - This is for using arena for creating new message instance.
* **reuse** - This is for reusing the same message instance for parsing.
* **Java** - For Java there're 3 kinds of parsing/Serialization ways:
* **byte[]** - This is for parsing from a Byte Array.
* **ByteString** - This is for parsing from a
com.google.protobuf.ByteString.
* **InputStream** - This is for parsing from a InputStream
* **Python** - For Pythong there're 3 kinds of python protobuf for testing:
* **C++-genereated-code** - This is for using cpp generated code of the
proto file as dynamic linked library.
* **C++-reflection** - This is for using cpp reflection, which there's no
generated code, but still using cpp protobuf library as dynamic linked
library.
* **pure-Python** - This is for pure Python version, which don't link with
any cpp protobuf library.
## Parsing performance
<table>
<tbody><tr>
<th rowspan="2"> </th>
<th colspan="3" rowspan="1">C++</th>
<th colspan="3" rowspan="1">C++ with tcmalloc</th>
<th colspan="3" rowspan="1">java</th>
<th colspan="3" rowspan="1">python</th>
</tr>
<tr>
<th colspan="1">new</th>
<th colspan="1">new arena</th>
<th colspan="1">reuse</th>
<th colspan="1">new</th>
<th colspan="1">new arena</th>
<th colspan="1">reuse</th>
<th colspan="1">byte[]</th>
<th colspan="1">ByteString</th>
<th colspan="1">InputStream</th>
<th colspan="1">C++-generated-code</th>
<th colspan="1">C++-reflection</th>
<th colspan="1">pure-Python</th>
</tr>
<tr>
<td>google_message1_proto2</td>
<td>368.717MB/s</td>
<td>261.847MB/s</td>
<td>799.403MB/s</td>
<td>645.183MB/s</td>
<td>441.023MB/s</td>
<td>1.122GB/s</td>
<td>425.437MB/s</td>
<td>425.937MB/s</td>
<td>251.018MB/s</td>
<td>82.8314MB/s</td>
<td>47.6763MB/s</td>
<td>3.76299MB/s</td>
</tr>
<tr>
<td>google_message1_proto3</td>
<td>294.517MB/s</td>
<td>229.116MB/s</td>
<td>469.982MB/s</td>
<td>434.510MB/s</td>
<td>394.701MB/s</td>
<td>591.931MB/s</td>
<td>357.597MB/s</td>
<td>378.568MB/s</td>
<td>221.676MB/s</td>
<td>82.0498MB/s</td>
<td>39.9467MB/s</td>
<td>3.77751MB/s</td>
</tr>
<tr>
<td>google_message2</td>
<td>277.242MB/s</td>
<td>347.611MB/s</td>
<td>793.67MB/s</td>
<td>503.721MB/s</td>
<td>596.333MB/s</td>
<td>922.533MB/s</td>
<td>416.778MB/s</td>
<td>419.543MB/s</td>
<td>367.145MB/s</td>
<td>241.46MB/s</td>
<td>71.5723MB/s</td>
<td>2.73538MB/s</td>
</tr>
<tr>
<td>google_message3_1</td>
<td>213.478MB/s</td>
<td>291.58MB/s</td>
<td>543.398MB/s</td>
<td>539.704MB/s</td>
<td>717.300MB/s</td>
<td>927.333MB/s</td>
<td>684.241MB/s</td>
<td>704.47MB/s</td>
<td>648.624MB/s</td>
<td>209.036MB/s</td>
<td>142.356MB/s</td>
<td>15.3324MB/s</td>
</tr>
<tr>
<td>google_message3_2</td>
<td>672.685MB/s</td>
<td>802.767MB/s</td>
<td>1.21505GB/s</td>
<td>985.790MB/s</td>
<td>1.136GB/s</td>
<td>1.367GB/s</td>
<td>1.54439GB/s</td>
<td>1.60603GB/s</td>
<td>1.33443GB/s</td>
<td>573.835MB/s</td>
<td>314.33MB/s</td>
<td>15.0169MB/s</td>
</tr>
<tr>
<td>google_message3_3</td>
<td>207.681MB/s</td>
<td>140.591MB/s</td>
<td>535.181MB/s</td>
<td>369.743MB/s</td>
<td>262.301MB/s</td>
<td>556.644MB/s</td>
<td>279.385MB/s</td>
<td>304.853MB/s</td>
<td>107.575MB/s</td>
<td>32.248MB/s</td>
<td>26.1431MB/s</td>
<td>2.63541MB/s</td>
</tr>
<tr>
<td>google_message3_4</td>
<td>7.96091GB/s</td>
<td>7.10024GB/s</td>
<td>9.3013GB/s</td>
<td>8.518GB/s</td>
<td>8.171GB/s</td>
<td>9.917GB/s</td>
<td>5.78006GB/s</td>
<td>5.85198GB/s</td>
<td>4.62609GB/s</td>
<td>2.49631GB/s</td>
<td>2.35442GB/s</td>
<td>802.061MB/s</td>
</tr>
<tr>
<td>google_message3_5</td>
<td>76.0072MB/s</td>
<td>51.6769MB/s</td>
<td>237.856MB/s</td>
<td>178.495MB/s</td>
<td>111.751MB/s</td>
<td>329.569MB/s</td>
<td>121.038MB/s</td>
<td>132.866MB/s</td>
<td>36.9197MB/s</td>
<td>10.3962MB/s</td>
<td>8.84659MB/s</td>
<td>1.25203MB/s</td>
</tr>
<tr>
<td>google_message4</td>
<td>331.46MB/s</td>
<td>404.862MB/s</td>
<td>427.99MB/s</td>
<td>589.887MB/s</td>
<td>720.367MB/s</td>
<td>705.373MB/s</td>
<td>606.228MB/s</td>
<td>589.13MB/s</td>
<td>530.692MB/s</td>
<td>305.543MB/s</td>
<td>174.834MB/s</td>
<td>7.86485MB/s</td>
</tr>
</tbody></table>
## Serialization performance
<table>
<tbody><tr>
<th rowspan="2"> </th>
<th colspan="1" rowspan="2">C++</th>
<th colspan="1" rowspan="2">C++ with tcmalloc</th>
<th colspan="3" rowspan="1">java</th>
<th colspan="3" rowspan="1">python</th>
</tr>
<tr>
<th colspan="1">byte[]</th>
<th colspan="1">ByteString</th>
<th colspan="1">InputStream</th>
<th colspan="1">C++-generated-code</th>
<th colspan="1">C++-reflection</th>
<th colspan="1">pure-Python</th>
</tr>
<tr>
<td>google_message1_proto2</td>
<td>1.39698GB/s</td>
<td>1.701GB/s</td>
<td>1.12915GB/s</td>
<td>1.13589GB/s</td>
<td>758.609MB/s</td>
<td>260.911MB/s</td>
<td>58.4815MB/s</td>
<td>5.77824MB/s</td>
</tr>
<tr>
<td>google_message1_proto3</td>
<td>959.305MB/s</td>
<td>939.404MB/s</td>
<td>1.15372GB/s</td>
<td>1.07824GB/s</td>
<td>802.337MB/s</td>
<td>239.4MB/s</td>
<td>33.6336MB/s</td>
<td>5.80524MB/s</td>
</tr>
<tr>
<td>google_message2</td>
<td>1.27429GB/s</td>
<td>1.402GB/s</td>
<td>1.01039GB/s</td>
<td>1022.99MB/s</td>
<td>798.736MB/s</td>
<td>996.755MB/s</td>
<td>57.9601MB/s</td>
<td>4.09246MB/s</td>
</tr>
<tr>
<td>google_message3_1</td>
<td>1.31916GB/s</td>
<td>2.049GB/s</td>
<td>991.496MB/s</td>
<td>860.332MB/s</td>
<td>662.88MB/s</td>
<td>1.48625GB/s</td>
<td>421.287MB/s</td>
<td>18.002MB/s</td>
</tr>
<tr>
<td>google_message3_2</td>
<td>2.15676GB/s</td>
<td>2.632GB/s</td>
<td>2.14736GB/s</td>
<td>2.08136GB/s</td>
<td>1.55997GB/s</td>
<td>2.39597GB/s</td>
<td>326.777MB/s</td>
<td>16.0527MB/s</td>
</tr>
<tr>
<td>google_message3_3</td>
<td>650.456MB/s</td>
<td>1.040GB/s</td>
<td>593.52MB/s</td>
<td>580.667MB/s</td>
<td>346.839MB/s</td>
<td>123.978MB/s</td>
<td>35.893MB/s</td>
<td>2.32834MB/s</td>
</tr>
<tr>
<td>google_message3_4</td>
<td>8.70154GB/s</td>
<td>9.825GB/s</td>
<td>5.88645GB/s</td>
<td>5.93946GB/s</td>
<td>2.44388GB/s</td>
<td>5.9241GB/s</td>
<td>4.05837GB/s</td>
<td>876.87MB/s</td>
</tr>
<tr>
<td>google_message3_5</td>
<td>246.33MB/s</td>
<td>443.993MB/s</td>
<td>283.278MB/s</td>
<td>259.167MB/s</td>
<td>206.37MB/s</td>
<td>37.0285MB/s</td>
<td>12.2228MB/s</td>
<td>1.1979MB/s</td>
</tr>
<tr>
<td>google_message4</td>
<td>1.56674GB/s</td>
<td>2.19601GB/s</td>
<td>776.907MB/s</td>
<td>770.707MB/s</td>
<td>702.931MB/s</td>
<td>1.49623GB/s</td>
<td>205.116MB/s</td>
<td>8.93428MB/s</td>
</tr>
</tbody></table>
\* The cpp performance can be improved by using [tcmalloc](https://gperftools.github.io/gperftools/tcmalloc.html), please follow the (instruction)[https://github.com/google/protobuf/blob/master/benchmarks/README.md] to link with tcmalloc to get the faster result.