[HDFS-13762] Support non-volatile storage class memory(SCM) in HDFS cache directives

• Type: New Feature
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.3.0
• Component/s: caching, datanode
• Target Version/s: 3.3.0
• Hadoop Flags: Reviewed

Release Note

Non-volatile storage class memory (SCM, also known as persistent memory) is supported in HDFS cache. To enable SCM cache, user just needs to configure SCM volume for property “dfs.datanode.cache.pmem.dirs” in hdfs-site.xml. And all HDFS cache directives keep unchanged. There are two implementations for HDFS SCM Cache, one is pure java code implementation and the other is native PMDK based implementation. The latter implementation can bring user better performance gain in cache write and cache read. If PMDK native libs could be loaded, it will use PMDK based implementation otherwise it will fallback to java code implementation. To enable PMDK based implementation, user should install PMDK library by referring to the official site http://pmem.io/. Then, build Hadoop with PMDK support by referring to "PMDK library build options" section in `BUILDING.txt` in the source code. If multiple SCM volumes are configured, a round-robin policy is used to select an available volume for caching a block. Consistent with DRAM cache, SCM cache also has no cache eviction mechanism. When DataNode receives a data read request from a client, if the corresponding block is cached into SCM, DataNode will instantiate an InputStream with the block location path on SCM (pure java implementation) or cache address on SCM (PMDK based implementation). Once the InputStream is created, DataNode will send the cached data to the client. Please refer "Centralized Cache Management" guide for more details.

• Assignee: Feilong He
• Reporter: Sammi Chen
• Created: 24/Jul/18 09:10
• Updated: 02/Jan/20 09:18
• Resolved: 15/Oct/19 07:04

Description

No-volatile storage class memory is a type of memory that can keep the data content after power failure or between the power cycle. Non-volatile storage class memory device usually has near access speed as memory DIMM while has lower cost than memory.  So today It is usually used as a supplement to memory to hold long tern persistent data, such as data in cache. 

Currently in HDFS, we have OS page cache backed read only cache and RAMDISK based lazy write cache.  Non-volatile memory suits for both these functions. 

This Jira aims to enable storage class memory first in read cache. Although storage class memory has non-volatile characteristics, to keep the same behavior as current read only cache, we don't use its persistent characteristics currently. 

 

 

 

Issue Links

• is related to: HDFS-14905 Backport HDFS persistent memory read cache support to branch-3.2
• relates to: HDFS-15080 Fix the issue in reading persistent memory cached data with an offset
• relates to: HDFS-14745 Backport HDFS persistent memory read cache support to branch-3.1
• relates to: HDFS-14740 Recover data blocks from persistent memory read cache during datanode restarts

Sub-Tasks

• Refactor MappableBlock to align with the implementation of SCM cache: Resolved
• Refactor FsDatasetCache for SCM cache implementation: Resolved
• Implement HDFS cache on SCM by using pure java mapped byte buffer: Resolved
• Refine the implementation for HDFS cache on SCM: Resolved
• Use FileChannel.transferTo() method for transferring block to SCM cache: Resolved
• Implement HDFS cache on SCM with native PMDK libs: Resolved
• Report pmem stats to namenode: Resolved
• Update documentation for HDFS cache on SCM support: Resolved
• Clean up pmem cache before setting pmem cache capacity: Resolved
• Check native pmdk lib by 'hadoop checknative' command: Resolved