Skip to content
/ SSM Public
forked from Intel-bigdata/SSM

Smart Storage Management for Big Data, a comprehensive hot/cold data optimized solution

License

Notifications You must be signed in to change notification settings

PHILO-HE/SSM

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HDFS Smart Storage Management Build Status

HDFS-SSM is the major portion of the overall Smart Data Management Initiative.

In big data field, HDFS storage has been facing increasing pressure due to various workloads and demanding performance in recent years. The latest storage devices (Optane Memory, Optane SSD, NVMe SSD, etc.) can be used to improve the storage performance. Meanwhile HDFS provides all kinds of nice methodologies like HDFS Cache, Heterogeneous Storage Management (HSM) and Erasure Coding (EC), but it is a big challenge for users to make full utilization of these high-performance storage devices and HDFS storage options in a dynamic environment.

To overcome the challenge, we have introduced a comprehensive end-to-end solution, aka Smart Storage Management (SSM) in Apache Hadoop. HDFS operation data and system state information are collected, and based on the collected metrics, SSM can automatically make sophisticated usage of these methodologies to optimize HDFS storage efficiency.

High Level Goals

1. Enhancement for HDFS-HSM and HDFS-Cache

Automatically and smartly adjusting storage policies and options according to data temperature. Already released.

2. Support block level erasure coding

Similar to the old HDFS-RAID, not only for Hadoop 3.x, but also for Hadoop 2.x. Ref. the block level erasure coding design.

3. Small files support and compaction

Optimizing NameNode to support even larger namespace, and eliminating the number of inodes for small files from memory. Supporting both write and read. Ref. the HDFS small files support design.

4. Cluster Data Copy and Disaster Recovery

Supporting transparent fail-over for applications. Here is the HDFS disaster recovery design document. The 1st stage already released.

5. Transparent HDFS Data Compression

Supporting transparent HDFS data compression. Note that it is not Hadoop compression, which needs to be explicitly called by applications and frameworks like MR.

High Level Considerations

  1. Supports Hadoop 3.x and Hadoop 2.x;
  2. The whole work and framework builds on top of HDFS, avoiding modifications when possible;
  3. Compatible HDFS client APIs to support existing applications and computation frameworks;
  4. Provide additional client APIs to allow new applications to benefit from SSM nice facilities;
  5. Support high availability and reliability, trying to reuse existing infrastructures in a deployment;
  6. Security is ensured, particularly when Hadoop security is enabled.

Architecture

SSM architecture is shown as follows. Ref. SSM design for details. Note that some of the contents need to be updated according to the latest implementation.

How SSM server and agents collaborate to perform one specific service? The below picture illustrates how SSM server and agents collaborate to perform some services.

The following picture depicts SSM system behaviours.

Below figure illustrates how to position SSM in big data ecosystem.

Development Phases

HDFS-SSM development is separated into 3 major phases. Currently the Phase 2 work is approaching completion.

Phase 1. Implement SSM framework and the fundamental infrastructure:

  • Event and metrics collection from HDFS cluster;
  • Rule DSL to support high level customization and usage;
  • Support various smart actions to adjust storage policies and options, enhancing HDFS-HSM and HDFS-Cache;
  • Client API, Admin API following Hadoop RPC and REST channels;
  • Basic web UI support.

Phase 2. Refine SSM framework and support user solutions:

  • Small files support and compaction;
  • Cluster disaster recovery;
  • Support block level erasure coding;
  • Transparent HDFS data compression;
  • To support the new desired actions, enhance the SSM framework and infrastructure.

Phase 3. Optimize further for computing frameworks and workloads benefiting from SSM offerings and facilities:

  • Hive on SSM;
  • HBase on SSM;
  • Spark on SSM;
  • Deep Learning on SSM.

Phase I -- Use Cases

1. Cache most hot data

When the files get very hot, they can be moved from fast storage into cache memory to achieve the best read performance. The following shows the example of moving data to memory cache if the data has been read over 3 times during the last 5 minutes

2. Move hot data to fast storage

Without SSM, data may always be read from HDD. With SSM, optimizations can be made through rules. As shown in the figure above, data can be moved to faster SSD to achieve better performance.

3. Archive cold data

Files are less likely to be read during the ending of lifecycle, so it is better to move these cold files into lower performance storage to decrease the cost of data storage. The following shows the example of archiving the data that has not been read during the last 90 days.

Admin Doc

Cluster administrator takes the role of SSM rule management. A set of APIs is exposed to help administrator manage rule. This set of APIs includes create, delete, list, enable and disable SSM rule. Hadoop admin privilege is required to access the APIs. For detailed information, please refer to Admin Guide.

User Doc

SSM provides a SmartDFSClient that includes both original HDFS DFSClient APIs and new SSM APIs. Applications can use this SmartDFSClient to benefit from the provided SSM facilities. New SSM APIs include cache file and archive file etc. More APIs will be added later. For detailed information, please refer to User Guide.

How to Build

Currently SSM supports CDH 5.10.1 (which is based on hadoop-2.6.0), hadoop-2.7.X and hadoop-3.1.X. So you can build SSM with the following commands accordingly.

mvn package -D skipTests -P dist,web,hadoop-cdh-2.6
mvn package -D skipTests -P dist,web,hadoop-2.7
mvn package -D skipTests -P dist,web,hadoop-3.1

Then a package named as smart-data-${version}.tar.gz will be generated under smart-dist/target.

Third-party dependencies

SSM uses SQL database as its backend metastore. MySQL is used in our multiple tests and it is recommended in production environment.

MySQL : User should deploy MySQL service in advance.

Developers

  • Wei, Zhou
  • Huafeng, Wang
  • Qiyuan, Gong
  • Yi, Chen
  • Kai, Zheng
  • Jie, Tao
  • Shunyang, Yao
  • Zhiqiang, Zhang
  • Feilong, He
  • Frank, Zeng
  • Tianlun, Zhang
  • Yuxuan, Pan
  • Jiajia, Li
  • Zhen, Du
  • Yao, Chen
  • Chen, Feng
  • Yanping, You
  • Tianlong, Xu
  • Anyang, Wang
  • Conghui, Li
  • Weijun, Pang

How to Contribute

We welcome your feedback and contributions. Please feel free to submit issues or PRs, and we'll respond soon. Note that the project is evolving very fast.

Acknowledgement

This work derives from the discussions in Apache Hadoop JIRA HDFS-7343. We want to thank not only all the team members of this project, but also all the contributors for their valuable ideas and feedback.

For any security concern, please visit https://01.org/security.

About

Smart Storage Management for Big Data, a comprehensive hot/cold data optimized solution

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Java 82.5%
  • JavaScript 6.8%
  • HTML 3.2%
  • Python 1.4%
  • Shell 1.3%
  • CSS 1.3%
  • Other 3.5%