Shulou(Shulou.com)06/02 Report--

1. Challenges in performance optimization

The following is the architecture of the entire PaaS platform

It mainly includes these subsystems:

Micro-service governance framework: provides a series of distributed / micro-service governance capabilities for applications, such as automatic registration, discovery, governance, isolation, call analysis, etc., shielding the complexity of distributed systems.

Application scheduling and resource management framework: through the lifecycle management automation from application modeling, orchestration and deployment to resource scheduling, self-scaling and monitoring self-healing.

Application development pipeline framework: to open up a series of CI/CD whole process automation from writing code submission to automatic compilation and packaging, continuous integration, and automatic deployment.

Cloud middleware services: database, big data, communication and application middleware services needed for cloud application; traditional non-cloud middleware capabilities can be integrated through service integration management and control.

In the face of such a complex system, performance optimization is a very difficult challenge, here are some pain points:

Many source code and development components, 100 + git repo, built more than 1 day as a whole

The running architecture is complex, and a full set of installation requires a 30-minute VMJ 200 + process.

The software stack is deep and the network plane is complex

The scale of the cluster is large, and it is very difficult to build a 5k-10k node environment.

The operation of the system will go through multiple distributed components, and the bottleneck of the system cannot be found through a single component diagnosis.

Unable to track the delay and throughput of thousands of API at different levels

Most developers focus on functional development and fail to realize that their code may cause performance problems

two。 Optimization analysis

So, how should we optimize such a large and complex system from a methodological point of view? The basic idea is to split a big problem into several uncoupled dimensions and break it down one by one. In terms of large dimensions, a PaaS container cluster can be divided into three large subsystems.

Control subsystem: control the issuance and operation of instructions (k8s), such as creating a pod

Business traffic subsystem: container network (flannel), load balancing (ELB/kube-proxy)

Monitoring subsystem: monitoring alarm data collection (kafka, Hadoop)

This seems to be just an architectural division, so how does it correspond to a specific business scenario? We can consider the following scenario in which applications are deployed in large quantities on the PaaS platform. See what pressure it will put on each subsystem in the process of deploying the application.

Application package size: 400m

Application template size: 10m

1000 nodes, one POD per node, one instance

10 types of software packages, depending on the length of 3pm 10GB network

Scheduling and Resource Management 3VM

These are some specifications for a typical deployment application, so for such an input, we can decompose the pressure on each subsystem according to the architecture, and the indicators that the subsystem needs to support are as follows:

Control subsystem: kubernetes scheduling speed > 50 pods/s, warehouse supports 300 concurrent downloads, > 40M/s

Data subsystem: overlay container network TCP transceiver performance loss

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.