Pangu — The High Performance Distributed File System by Alibaba Cloud

Pangu is a high-reliability, high-availability, and high-performance distributed file system developed by Alibaba Cloud. It has a history of nearly ten years. As a unified storage core of Alibaba Cloud, Pangu 1.0 stably and efficiently supported the rapid development of multiple business lines of Alibaba Cloud, including ECS, NAS, OSS, Table Store, MaxCompute, and AnalyticDB. In recent years, Alibaba Cloud redesigned and implemented the second-generation storage engine Pangu 2.0. It provides better storage services not only for Alibaba Cloud but also for businesses of Alibaba Group and Ant Financial. There are two reasons for the upgrade: hardware improvements and demands from the business.

Rapid Development of Underlying Hardware

In recent years, the performance of distributed storage-related hardware has dramatically improved.

Pressure from the Upper-Layer Business

Currently, an increasing number of businesses are getting connected to Pangu. However, as enterprises have different requirements for storage, adequately supporting businesses has become an enormous pressure for Pangu.

Design Objectives

To cope with the above-mentioned new trends and challenges, Pangu 2.0 has set up the following core design objectives:

  1. Fully Distributed Metadata Management: Performs fully distributed management and dynamic splitting and migration of metadata to greatly increase the number of managed files, resolve the problem of special model dependence of metadata nodes, reduce the “explosion radius” of faults, and improve the platform stability.
  2. System Elasticity: Supports multiple product forms and shares the core paths to provide the scalability for access of more businesses in the future and prevent architecture adjustment caused by business access; unifies the hardware access interfaces and supports access of current and future new hardware optimally.
  3. Optimized Cost: Adopts the hierarchy, erasure coding (EC), compression, and deduplication technologies to reduce the storage cost, wins the initiative in increasingly fierce business competition, and gains technical advantages to cope with the exponential data growth.

Pangu 2.0’s Architecture

To achieve the preceding objectives, Pangu 2.0 adopts the hierarchical architecture, as shown in the following figure:

Core Base

The core base layer of Pangu consists of the client, MetaServer, and ChunkServer.

Fully Distributed Metadata Management

Pangu 2.0 adopts a fully distributed metadata management policy, in which MetaServer is horizontally expandable. Data partitioning is done based on the preset policy. Multiple MetaServers in each partition use the RAFT protocol to achieve high reliability and availability, and data in each MetaServer is persistent in ChunkServer in metachunk mode.

Efficient I/O Path

Fully distributed metadata management ensures higher stability and scalability of Pangu, and the efficient I/O path resolves the performance problem of Pangu.

Excellent Thread Model

The architecture of Pangu 2.0 has several advantages. For engineering implementation and better performance, the data path of Pangu uses the run-to-completion concurrency model. The entire I/O request gets processed in one thread, saving the overheads for thread synchronization, CPU cache miss, and context switch. To obtain the optimal performance using this concurrency model, Pangu well designs the underlying RDMA/TCP network library, SPDK I/O library, and modules at the business layer and connects all links.

High-Performance Network Library

Pangu has minimized its performance overheads because the hardware performance must be flexible and efficiently exploited to create the optimal-performance storage system. In an environment with the high-speed network, Pangu supports RPC communications using RDMA, which is encapsulated in the RPC and transparent to upper-layer businesses. Pangu can also flexibly implement TCP and RDMA communications based on requirements of different businesses.

Cost Control

The high-reliability and high-performance distributed file system ensure the position of Pangu in the industry, while effective cost control further brings the initiative to the business. Therefore, Pangu has made every effort to reduce the cost of using various policies.

  1. EC: The latest HDFS 3.0 supports EC, while Pangu 1.0 has supported backend EC and Pangu 2.0 has supported frontend EC. Compared with the multi-replica service, EC greatly reduces the I/O and network traffic. In some scenarios, EC also reduces the cost and increases the throughput.


In Alibaba Cloud, many users buy ECS virtual machines to build a Hadoop ecosystem for big data analysis, in which HDFS is a necessary component. However, as a non-cloud storage system, HDFS also has various disadvantages, such as poor elasticity, high management cost, the poor performance of small files, and lack of enterprise-level disaster recovery. Moreover, cloud users require HDFS to interconnect with other cloud data.

  1. MapReduce without YARN + DFS
  2. MapReduce with YARN + DFS
  3. Hive without YARN + DFS
  4. Hive with YARN + DFS
  5. Spark without YARN + DFS
  6. Spark with YARN + DFS
  7. TPC-DS test SparkSQL + DFS
  8. TPC-DS test Impala + DFS

Future Prospects

Although Pangu has reached the leading position in the industry, new technologies and requirements are emerging. If Pangu does not catch up with other competitors, it may be soon thrown out. Therefore, Pangu always keeps its eye on the evolution of the underlying depended hardware and upper-layer business features. For example, Pangu is trying to replace the device-based SSD with the host-based SSD to better utilize the hardware performance by integrating the software and hardware. Also, it is developing the full-link QoS function to resolve the problem of resource competition between multiple users or tasks in scenarios where storage and computing services are separated.

Follow me to keep abreast with the latest technology news, industry insights, and developer trends. Alibaba Cloud website: