Support specifying storage type for per-application local dirs

Introduction

• Some applications of various frameworks (Flink, Spark and MapReduce etc) using local storage (checkpoint, shuffle etc) might require high IO performance. It's useful to allocate local directories to high performance storage media for these applications on heterogeneous clusters.
• YARN does not distinguish different storage types and hence applications cannot selectively use storage media with different performance characteristics. Adding awareness of storage media can allow YARN to make better decisions about the placement of local directories.

Approach

• NodeManager will distinguish storage types for local directories.
  • yarn.nodemanager.local-dirs and yarn.nodemanager.log-dirs configuration should allow the cluster administrator to optionally specify the storage type for each local directories. Example: [SSD]/disk1/nm-local-dir,/disk2/nm-local-dir,/disk3/nm-local-dir (equals to [SSD]/disk1/nm-local-dir,[DISK]/disk2/nm-local-dir,[DISK]/disk3/nm-local-dir)
  • StorageType defines DISK/SSD storage types and takes DISK as the default storage type.
  • StorageLocation separates storage type and directory path, used by LocalDirAllocator to aware the types of local dirs, the default storage type is DISK.
  • getLocalPathForWrite method of LocalDirAllcator will prefer to choose the local directory of the specified storage type, and will fallback to not care storage type if the requirement can not be satisfied.
  • Support for container related local/log directories by ContainerLaunch. All application frameworks can set the environment variables (LOCAL_STORAGE_TYPE and LOG_STORAGE_TYPE) to specified the desired storage type of local/log directories, and choose to not launch container if fallback through these environment variables (ENSURE_LOCAL_STORAGE_TYPE and ENSURE_LOG_STORAGE_TYPE).
• Allow specified storage type for various frameworks (Take MapReduce as an example)
  • Add new configurations should allow application administrator to optionally specify the storage type of local/log directories and fallback strategy (MapReduce configurations: mapreduce.job.local-storage-type, mapreduce.job.log-storage-type, mapreduce.job.ensure-local-storage-type and mapreduce.job.ensure-log-storage-type).
  • Support for container work directories. Set the environment variables includes LOCAL_STORAGE_TYPE and LOG_STORAGE_TYPE according to configurations above for ContainerLaunchContext and ApplicationSubmissionContext. (MapReduce should update YARNRunner and TaskAttemptImpl)
  • Add storage type prefix for request path to support for other local directories of frameworks (such as shuffle directories for MapReduce). (MapReduce should update YarnOutputFiles, MROutputFiles and YarnChild to support for output/work directories)
  • Flow diagram for MapReduce framework
    

Further Discussion

• Scheduling : The requirement of storage type for local/log directories may not be satisfied for a part of nodes on heterogeneous clusters. To achieve global optimum, scheduler should aware and manage disk resources.
  • Approach-1: Based on node attributes (YARN-3409), Scheduler can allocate containers which have SSD requirement on nodes with attribute:ssd=true.
  • Approach-2: Based on extended resource model (YARN-3926), it's easy to support scheduling through extending resource models like vdisk and vssd using this feature, but hard to measure for applications and isolate for non-CFQ based disks.
• Fallback strategy still needs to be concerned. Certain applications might not work well when the requirement of storage type is not satisfied. When none of desired storage type disk are available, should container launching be failed? let AM handle? We have implemented a fallback strategy that fail to launch container when none of desired storage type disk are available. Is there some better methods?

This feature has been used for half a year to meet the needs of some applications on Alibaba search clusters.
Please feel free to give your suggestions and opinions.