Practice of RTC performance automation tool in memory optimization scenario

Background introduction

The performance test is  SDK  Important basis for publishing ,VolcRTC  All business parties pay more attention to performance indicators , about  RTC  There are clear access standards . Therefore, we have established an offline performance automated test system , During the test, we found that  VolcRTC  The memory occupation of is high, and there is a large optimization space . A version  1v1  voice call  VolcRTC 1v1  Voice call memory occupation ：

To achieve memory optimization , First of all, we need to clarify two issues ：

• Which modules consume how much memory ？

• How to optimize ？

The memory of

Before answering the above two questions , Let's first understand the main components of memory .

stay  Android  On the system , Memory is mainly divided into :

The red box in the following figure is  VolcRTC  During the call

• Java Heap,
  from  Java  Code assigned objects ; During the call  Java  Memory distribution curve , It mainly presents jagged periodic changes . combination  VolcRTC  Business characteristics , You can know that this part of memory is mainly  JNI  Allocate temporary objects when calling , After accumulating to a certain extent, the system  GC  Mechanism recycling .

• Native Heap,
  from  C  or  C++  Code assigned objects . This part is  VolcRTC  Main memory occupation .

• Code,
  For processing code and resources （ Such as  dex  Bytecode 、 Optimized or compiled  dex  Code 、.so  Library and Fonts ） Of memory .VolcRTC  Memory occupied by the Library , But it is not equal to the package size of the dynamic library , The main reason is that code segments are loaded on demand , So part of the code will not be loaded into memory .VolcRTC  It's a dynamic library , therefore  Code  Memory is also a major part of the call process .

To optimize the direction

According to the preliminary analysis above , Can be determined  VolcRTC  The memory occupation of is mainly distributed in  Native Heap  And  Code  paragraph . Therefore, we make it clear that the general optimization direction is ：

• Native  Memory optimization

• Dynamic library enclosure optimization

Memory attribution analysis

Which modules consume how much memory ？

• Memory allocation stack information

• Attribution by module

Heapprofd  Realization principle

• hook malloc、calloc、realloc、free  And other functions related to memory allocation

• Copy registers and stack memory , Stored in shared memory , For stack backtracking

• Cluster and generate according to stack information  Trace  file

Module attribution

VolcRTC  Attribution rules

VolcRTC  It is mainly divided into the bottom media engine and the top  RTC SDK  Two parts . The overall architecture of the media engine is based on pipeline （Pipeline） In the form of , Every  Pipeline  By realizing different functions  Node  constitute . We can filter the stack according to the relevant namespace , Then, according to the software layered architecture, attribution is carried out layer by layer .

Pure system stack

VolcRTC  System stack memory allocation caused , The stack does not contain  VolcRTC  Symbol information , Cannot be classified according to the above rules , Need to be classified into by  VolcRTC  Caused by system memory allocation .

Attribution examples

The memory allocation stack is generally characterized by

__pthread_start(void*)

, The top of the stack is the memory allocation method , In the middle  VolcRTC  Stack information . According to the stack information , Combine the attribution rules and then attribute upward layer by layer , Form a tree structure , Accurately analyze each  Pipeline、 every last  Node、 The amount of memory occupied by each type of object .

Problems encountered

Hook malloc  The resulting memory size and  Native Heap  Inconsistent size

malloc  Memory requested from memory allocator , It is similar to the one passed in when the program is running  size  Agreement .

The memory requested by the memory allocator to the operating system is allocated by page , Generally, each page is  4K,Native Heap  The statistics is the memory size of this part .

As the memory allocator needs to allocate and release memory repeatedly , Inevitably, there will be memory gaps, that is, memory fragments , In addition, the memory allocator will cache some small memory blocks to improve the efficiency of memory allocation .

Voice call memory analysis

Through performance automation testing tools , Generate analysis report . Based on the analysis report, we draw the panorama of voice call memory , Then identify several modules with high memory consumption through the panorama , Guide the optimization direction .

Memory optimization

Compile optimization

The packet size will directly affect the memory size , Therefore, optimizing the package size can also effectively reduce the memory size . By opening the  LTO、Oz  And other compilation options , Determine the compilation optimization options that need to be enabled in combination with the offline performance automation test to evaluate whether there is a negative impact on the performance indicators . After compiling and optimizing  Android  End dynamic library enclosures are reduced  900KB, Call process memory optimization  850KB  about .

Dynamic distribution on demand

VolcRTC  As a general function  SDK, There are many redundant logic and functions for each specific scenario , These logic and functions have more pre allocated memory . The corresponding optimization scheme is ：

• Reasonable code componentization , Abstract different functions into components to achieve flexible assembly and on-demand loading , Such as ：AI  The built-in data and models of noise reduction function will occupy a large memory space .

• Try to allocate memory dynamically on demand . Such as ：AEC  Echo cancellation has different algorithms in different scenes , You need to allocate memory as needed according to the actual scene , Reduce excessive memory pre allocation .

Set a reasonable cache size

Unreasonable cache size will also cause unnecessary memory waste . Through the attribution analysis of memory , Combine the business characteristics of different scenarios , Set a more reasonable cache size , It can reduce the memory consumption . for example ：RTC  Adopted  RTP  Packet retransmission mechanism to combat network packet loss , In order to implement the retransmission mechanism , You need to cache a certain number of packages , The number of caches needs to follow the frame length 、 Real time requirements and other business characteristics to set a reasonable value .

Reasonable algorithm and data structure design

Reasonable algorithm and data structure can also effectively reduce memory . On the premise of ensuring the accuracy of calculation , By reducing the data range , Use data types that occupy less memory space to implement the algorithm , For example, relative time rather than absolute time is used in statistics of time-related data 、 Spatial audio algorithm is used through fixed-point

short

Type instead of floating point data . In addition, memory alignment needs to be considered when designing data structures .

The optimization effect

1v1  voice call

This memory optimization , We explored  RTC  Performance attribution analysis drives performance optimization practice in scenarios . The following experience can be summed up ：

• Construct stable test cases

• Establish the data attribution model of performance impairment

• Identify hot performance issues based on attribution models , Form an optimization scheme

• from  1v1  Call start analysis , Then gradually to many people 、 Hundreds of people, thousands of people .

Join us

Bytes to beat  RTC  The team , As the world's leading audio and video team , We are committed to providing high-quality products on the global Internet 、 Low delay real-time audio and video communication capability , At present, it has supported Tiktok 、TikTok、 Qingbei online school 、 Byte game , Video conferencing and other applications , service  10  Billion users , Users are all over the world .

We are  

RTC Client（ client ） department ,

Dedicated to maintaining multiple functions / Platform  SDK  And best practices , Meet the requirements of flexibility and ease of use , In audio and video communication 、 Industry Applications 、 Signaling and other directions form their own characteristics , Look forward to the participation of excellent students .

Job list ：

https://bytedance.feishu.cn/docx/doxcnekgHRNsi2HO7LJIirFpw2e

Reference documents ：

Heapprofd: 

https://perfetto.dev/docs/desig