While the cluster fails a write request in a cluster that does not have at least k+p Swarm nodes available, the minimum number of required nodes is actually higher, to allow space for recovery in the event of a failed volume.
The minimum number of nodes to guarantee base functionality is k+(p*2).
The recommended number for performance and load distribution is (k+p)*2 .
5:2 encoding requires at least 9 Swarm nodes, preferably 14 nodes.
Nodes using erasure coding benefit from additional CPU cores (and Swarm multi-server processes to utilize them).
The recommended minimum of RAM per node is 4 GB.
Do not run Swarm using an EC level with fewer than 2 parity segments, except in cases of transient data or data that can be readily reproduced or regenerated.
Good baseline encoding is 4:2 or 5:2; this guarantees data protection equal to or better than replication, with a smaller footprint.
Subclusters have specific encoding considerations.
Write throughput for large objects that are erasure-coded is approximately equivalent to replicated objects written with Replicate on Write. Smaller objects are less performant due to the overhead of creating the parity segments relative to the time to write only the content body.
Parity Count. The biggest impact on performance for similar-sized objects is the parity count: the smaller the number of required parity segments, the faster the write performance.
Segment Size. If you can predict the size of the objects written to the cluster, you can increase the default segment size to reduce the number of EC sets it takes to encode the object, thereby increasing performance. Increasing the segment size of a 2 GB object with 5:2 encoding to 400 MB guarantees the object fits in one EC set.
Write Threads. Erasure-coded objects require more Swarm nodes participate in each write request. The number of client threads each node can service at a time is less than with replicated objects. For planning purposes, assume that each Swarm node can service 4 erasure-coding client write threads at a time before the cluster begins to return 503 (Service Unavailable) responses, indicating the cluster is too busy to service more requests. That threshold may differ depending on the size of the cluster and the encoding you use.