奇虎360和Go语言

jopen 9年前

奇虎 360 的软件工程师在 Go 语言官方博客上发表文章介绍该公司使用 Go 语言的经验

尽管 Google 被封杀,但该公司主导开发的 Go 语言在中国非常受欢迎,而让很多人感到不可思议的是,奇虎是一家受争议的公司,同时还是封杀 Google 的帮凶。

在 Go 博客文章中,奇虎的工程师称他们从 2012 年开始使用 Go,被用于旗下产品的推送服务。奇虎的产品多为免费,它的收入来源与推送关系密切。奇虎工程师称,基于 Go 的推送系统最初运行在 20 台服务器上,有 2000 万的实时连接,一天发送 200 万条信息。今天,这套系统运行在了 400 台服务器上,支持 2 亿+的实时连接,每天发送超过 100 亿条信息。

Qihoo 360 and Go

This guest blog post was written by Yang Zhou, Software Engineer at Qihoo 360.

Qihoo 360 is a major provider of Internet and mobile security products and services in China, and operates a major Android-based mobile distribution platform. At the end of June 2014, Qihoo had about 500 million monthly active PC Internet users and over 640 million mobile users. Qihoo also operates one of China’s most popular Internet browsers and PC search engines.

My team, the Push Service Team, provides fundamental messaging services for more than 50 products across the company (both PC and mobile), including thousands of Apps in our open platform.

Our "love affair" with Go dates back to 2012 when we first attempted to provide push services for one of Qihoo’s products. The initial version was built with nginx + lua + redis, which failed to satisfy our requirement for real-time performance due to excessive load. Under these circumstances, the newly-published Go 1.0.3 release came to our attention. We completed a prototype in a matter of weeks, largely thanks to the goroutine and channel features it provided.

Initially, our Go-based system ran on 20 servers, with 20 million real-time connections in total. The system sent 2 million messages a day. That system now runs on 400 servers, supporting 200 million+ real-time connections. It now sends over 10 billion messages daily.

With rapid business expansion and increasing application needs for our push service, the initial Go system quickly reached its bottleneck: heap size went up to 69G, with maximum garbage collection (GC) pauses of 3-6 seconds. Worse still, we had to reboot the system every week to release memory. It wouldn’t be honest if we didn’t consider relinquishing Go and instead, re-writing the entire core component with C. However, things didn’t go exactly as we planned, we ran into trouble migrating the code of Business Logic Layer. As a result, it was impossible for the only personnel at that time (myself) to maintain the Go system while ensuring the logic transfer to the C service framework.

Therefore, I made the decision to stay with Go system (probably the wisest one I had to make), and great headway was made soon enough.

Here are a few tweaks we made and key take-aways:

  • Replace short connections with persistent ones (using a connection pool), to reduce creation of buffers and objects during communication.
  • Use Objects and Memory pools appropriately, to reduce the load on the GC.

奇虎360和Go语言

  • Use a Task Pool, a mechanism with a group of long-lived goroutines consuming global task or message queues sent by connection goroutines, to replace short-lived goroutines.
  • Monitor and control goroutine numbers in the program. The lack of control can cause unbearable burden on the GC, imposed by surges in goroutines due to uninhibited acceptance of external requests, as RPC invocations sent to inner servers may block goroutines recently created.
  • Remember to add read and write deadlines to connections when under a mobile network; otherwise, it may lead to goroutine blockage. Apply it properly and with caution when under a LAN network, otherwise your RPC communication efficiency will be hurt.
  • Use Pipeline (under Full Duplex feature of TCP) to enhance the communication efficiency of RPC framework.

As a result, we successfully launched three iterations of our architecture, and two iterations of our RPC framework even with limited human resources. This can all attributed to the development convenience of Go. Below you can find the up-to-date system architecture:

奇虎360和Go语言

The continuous improvement journey can be illustrated by a table:

奇虎360和Go语言

Also, no temporary release of memory or system reboot is required after these optimizations.

What’s more exciting is we developed an on-line real-time Visibility Platform for profiling Go programs. We can now easily access and diagnose the system status, pinning down any potential risks. Here is a screen shot of the system in action:

奇虎360和Go语言

奇虎360和Go语言

The great thing about this platform is that we can actually simulate the connection and behavior of millions of online users, by applying the Distributed Stress Test Tool (also built using Go), and observe all real-time visualized data. This allows us to evaluate the effectiveness of any optimization and preclude problems by identifying system bottlenecks.

Almost every possible system optimization has been practiced so far. And we look forward to more good news from the GC team so that we could be further relieved from heavy development work. I guess our experience may also grow obsolete one day, as Go continues to evolve.

This is why I want to conclude my sharing by extending my sincere appreciation to the opportunity to attendGopher China. It was a gala for us to learn, to share and for offering a window showcasing Go’s popularity and prosperity in China. Many other teams within Qihoo have already either got to know Go, or tried to use Go.

I am convinced that many more Chinese Internet firms will join us in re-creating their system in Go and the Go team's efforts will benefit more developers and enterprises in the foreseeable future.

By Yang Zhou

来自: Solidot