HDFS Erasure Coding

Purpose

Replication is expensive – the default 3x replication scheme in HDFS has 200% overhead in storage space and other resources (e.g., network bandwidth). However, for warm and cold datasets with relatively low I/O activities, additional block replicas are rarely accessed during normal operations, but still consume the same amount of resources as the first replica.

Therefore, a natural improvement is to use Erasure Coding (EC) in place of replication, which provides the same level of fault-tolerance with much less storage space. In typical Erasure Coding (EC) setups, the storage overhead is no more than 50%. Replication factor of an EC file is meaningless. It is always 1 and cannot be changed via -setrep command.

Background

In storage systems, the most notable usage of EC is Redundant Array of Inexpensive Disks (RAID). RAID implements EC through striping, which divides logically sequential data (such as a file) into smaller units (such as bit, byte, or block) and stores consecutive units on different disks. In the rest of this guide this unit of striping distribution is termed a striping cell (or cell). For each stripe of original data cells, a certain number of parity cells are calculated and stored – the process of which is called encoding. The error on any striping cell can be recovered through decoding calculation based on surviving data and parity cells.

Integrating EC with HDFS can improve storage efficiency while still providing similar data durability as traditional replication-based HDFS deployments. As an example, a 3x replicated file with 6 blocks will consume 6*3 = 18 blocks of disk space. But with EC (6 data, 3 parity) deployment, it will only consume 9 blocks of disk space.

Architecture

In the context of EC, striping has several critical advantages. First, it enables online EC (writing data immediately in EC format), avoiding a conversion phase and immediately saving storage space. Online EC also enhances sequential I/O performance by leveraging multiple disk spindles in parallel; this is especially desirable in clusters with high end networking. Second, it naturally distributes a small file to multiple DataNodes and eliminates the need to bundle multiple files into a single coding group. This greatly simplifies file operations such as deletion, quota reporting, and migration between federated namespaces.

In typical HDFS clusters, small files can account for over 3/4 of total storage consumption. To better support small files, in this first phase of work HDFS supports EC with striping. In the future, HDFS will also support a contiguous EC layout. See the design doc and discussion on HDFS-7285 for more information.

• NameNode Extensions - Striped HDFS files are logically composed of block groups, each of which contains a certain number of internal blocks. To reduce NameNode memory consumption from these additional blocks, a new hierarchical block naming protocol was introduced. The ID of a block group can be inferred from the ID of any of its internal blocks. This allows management at the level of the block group rather than the block.

• Client Extensions - The client read and write paths were enhanced to work on multiple internal blocks in a block group in parallel. On the output / write path, DFSStripedOutputStream manages a set of data streamers, one for each DataNode storing an internal block in the current block group. The streamers mostly work asynchronously. A coordinator takes charge of operations on the entire block group, including ending the current block group, allocating a new block group, and so forth. On the input / read path, DFSStripedInputStream translates a requested logical byte range of data as ranges into internal blocks stored on DataNodes. It then issues read requests in parallel. Upon failures, it issues additional read requests for decoding.

• DataNode Extensions - The DataNode runs an additional ErasureCodingWorker (ECWorker) task for background recovery of failed erasure coded blocks. Failed EC blocks are detected by the NameNode, which then chooses a DataNode to do the recovery work. The recovery task is passed as a heartbeat response. This process is similar to how replicated blocks are re-replicated on failure. Reconstruction performs three key tasks:

  1. Read the data from source nodes: Input data is read in parallel from source nodes using a dedicated thread pool. Based on the EC policy, it schedules the read requests to all source targets and reads only the minimum number of input blocks for reconstruction.

  2. Decode the data and generate the output data: New data and parity blocks are decoded from the input data. All missing data and parity blocks are decoded together.

  3. Transfer the generated data blocks to target nodes: Once decoding is finished, the recovered blocks are transferred to target DataNodes.

• Erasure coding policies To accommodate heterogeneous workloads, we allow files and directories in an HDFS cluster to have different replication and erasure coding policies. The erasure coding policy encapsulates how to encode/decode a file. Each policy is defined by the following pieces of information:

  1. The EC schema: This includes the numbers of data and parity blocks in an EC group (e.g., 6+3), as well as the codec algorithm (e.g., Reed-Solomon, XOR).

  2. The size of a striping cell. This determines the granularity of striped reads and writes, including buffer sizes and encoding work.

  Policies are named codec-num data blocks-num parity blocks-cell size. Currently, six built-in policies are supported: RS-3-2-1024k, RS-6-3-1024k, RS-10-4-1024k, RS-LEGACY-6-3-1024k, XOR-2-1-1024k and REPLICATION.

  REPLICATION is a special policy. It can only be set on directory, to force the directory to adopt 3x replication scheme, instead of inheriting its ancestor’s erasure coding policy. This policy makes it possible to interleave 3x replication scheme directory with erasure coding directory.

  REPLICATION policy is always enabled. For other built-in policies, they are disabled by default.

  Similar to HDFS storage policies, erasure coding policies are set on a directory. When a file is created, it inherits the EC policy of its nearest ancestor directory.

  Directory-level EC policies only affect new files created within the directory. Once a file has been created, its erasure coding policy can be queried but not changed. If an erasure coded file is renamed to a directory with a different EC policy, the file retains its existing EC policy. Converting a file to a different EC policy requires rewriting its data; do this by copying the file (e.g. via distcp) rather than renaming it.

  We allow users to define their own EC policies via an XML file, which must have the following three parts:

  1. layoutversion: This indicates the version of EC policy XML file format.

  2. schemas: This includes all the user defined EC schemas.

  3. policies: This includes all the user defined EC policies, and each policy consists of schema id and the size of a striping cell (cellsize).

  A sample EC policy XML file named user_ec_policies.xml.template is in the Hadoop conf directory, which user can reference.

• Intel ISA-L Intel ISA-L stands for Intel Intelligent Storage Acceleration Library. ISA-L is an open-source collection of optimized low-level functions designed for storage applications. It includes fast block Reed-Solomon type erasure codes optimized for Intel AVX and AVX2 instruction sets. HDFS erasure coding can leverage ISA-L to accelerate encoding and decoding calculation. ISA-L supports most major operating systems, including Linux and Windows. ISA-L is not enabled by default. See the instructions below for how to enable ISA-L.

Deployment

   hdfs ec [generic options]
     [-setPolicy -path <path> [-policy <policyName>] [-replicate]]
     [-getPolicy -path <path>]
     [-unsetPolicy -path <path>]
     [-listPolicies]
     [-addPolicies -policyFile <file>]
     [-listCodecs]
     [-enablePolicy -policy <policyName>]
     [-disablePolicy -policy <policyName>]
     [-help [cmd ...]]

Limitations

Certain HDFS file write operations, i.e., hflush, hsync and append, are not supported on erasure coded files due to substantial technical challenges.

• append() on an erasure coded file will throw IOException.
• hflush() and hsync() on DFSStripedOutputStream are no-op. Thus calling hflush() or hsync() on an erasure coded file can not guarantee data being persistent.

A client can use StreamCapabilities API to query whether a OutputStream supports hflush() and hsync(). If the client desires data persistence via hflush() and hsync(), the current remedy is creating such files as regular 3x replication files in a non-erasure-coded directory, or using FSDataOutputStreamBuilder#replicate() API to create 3x replication files in an erasure-coded directory.