Erasure Coding EC

Replication is a proven and valuable mechanism to verify data integrity, but the cost per GB of storage can become high as object sizes and cluster sizes expand. A complementary data protection strategy, erasure coding (EC), provides high data durability with a smaller footprint. Swarm manages EC and replication together to optimize cost-effectiveness, converting objects between them seamlessly and dynamically, based on the policies set.

How EC works

Erasure coding breaks the original object into multiple data segments (k) and computes additional parity segments (p) based on the content of the data segments. This results in m total segments (k + p = m) being distributed to m different nodes (or subclusters) in the storage cluster (see the ec.protectionLevel setting).

The erasure coding encoding level is expressed as a tuple in this format:

{data segments}:{parity segments}

Swarm creates multiple sets of erasure segments for very large objects. The object breakdown into one or more erasure sets is transparent to external applications. A GET or HEAD of an erasure-coded object uses the same syntax as a replicated object.

The following illustration represents how erasure coding works:

How much EC protects

Swarm still reads the object as long as there are still k total segments (any combination of original data or parity) remaining in the cluster if a hard disk or a node containing an erasure segment fails. Protection against disk failure for the object is equal to the number of specified parity p segments.

The segments from a 5:2 (5 data segments with 2 parity segments for a total of 7 segments) or 8:2 (8 data and 2 parity segments for 10 total segments) erasure code are protected against the loss of any two nodes because they are distributed to different nodes. An erasure-coded object is immediately retrievable when accessed even if some segments are missing. Regenerating the missing erasure set segments is performed in a self-healing, cluster-initiated manner (similar to the recovery process for replicated objects) to protect against further disk loss. This process kicks off automatically when a missing volume is detected and automatically regenerates any missing segments.

Swarm always seeks the widest possible distribution of content (whether replicas or EC segments) regardless of settings, to maximize data protection. Swarm applies the following 2p+2 protection levels as possible:

1. One segment per subcluster … up to p segments per subcluster

2. One segment per node … up to p segments per node

3. One segment per volume

How much disk space EC saves

The amount of disk space (or footprint) used for erasure-coded objects depends on the ratio of data to parity segments in the specified encoding.

Use the following formula to roughly calculate the disk space expected to be used by an EC object with one set of erasure segments:

How footprint changes with different EC encoding (versus 3 reps)

• 1 GB object with 5:2 encoding: ((5 + 2) ÷ 5) × 1 GB = 1.4 GB (vs. 3 GB for replication)

• 3 GB object with 5:2 encoding: ((5 + 2) ÷ 5) × 3 GB = 4.2 GB (vs. 9 GB for replication)

• 3 GB object with 7:3 encoding: ((7 + 3) ÷ 7) × 3 GB = 4.3 GB (vs. 9 GB for replication)

See Hardware Requirements for Storage for how to size and optimize hardware for erasure coding.

See Cluster Protection Planning for how to determine the required number of nodes and optimize EC performance in a cluster.