CLAY code plugin

CLAY (short for coupled-layer) codes are erasure codes designed to bring about significant savings in terms of network bandwidth and disk IO when a failed node/OSD/rack is being repaired. Let:

d = number of OSDs contacted during repair

If jerasure is configured with k=8 and m=4, losing one OSD requires reading from the d=8 others to repair. And recovery of say a 1GiB needs a download of 8 X 1GiB = 8GiB of information.

However, in the case of the clay plugin d is configurable within the limits:

k+1 <= d <= k+m-1

By default, the clay code plugin picks d=k+m-1 as it provides the greatest savings in terms of network bandwidth and disk IO. In the case of the clay plugin configured with k=8, m=4 and d=11 when a single OSD fails, d=11 osds are contacted and 250MiB is downloaded from each of them, resulting in a total download of 11 X 250MiB = 2.75GiB amount of information. More general parameters are provided below. The benefits are substantial when the repair is carried out for a rack that stores information on the order of Terabytes.

plugin total amount of disk IO
jerasure,isa k*S
clay d*S/(d-k+1) = (k+m-1)*S/m

where S is the amount of data stored on a single OSD undergoing repair. In the table above, we have used the largest possible value of d as this will result in the smallest amount of data download needed to achieve recovery from an OSD failure.

Erasure-code profile examples

An example configuration that can be used to observe reduced bandwidth usage:

$ ceph osd erasure-code-profile set CLAYprofile \
     plugin=clay \
     k=4 m=2 d=5 \
     crush-failure-domain=host
$ ceph osd pool create claypool 12 12 erasure CLAYprofile

Creating a clay profile

To create a new clay code profile:

ceph osd erasure-code-profile set {name} \
     plugin=clay \
     k={data-chunks} \
     m={coding-chunks} \
     [d={helper-chunks}] \
     [scalar_mds={plugin-name}] \
     [technique={technique-name}] \
     [crush-failure-domain={bucket-type}] \
     [directory={directory}] \
     [--force]

Where:

k={data chunks}

Description:Each object is split into data-chunks parts, each of which is stored on a different OSD.
Type:Integer
Required:Yes.
Example:4

m={coding-chunks}

Description:Compute coding chunks for each object and store them on different OSDs. The number of coding chunks is also the number of OSDs that can be down without losing data.
Type:Integer
Required:Yes.
Example:2

d={helper-chunks}

Description:Number of OSDs requested to send data during recovery of a single chunk. d needs to be chosen such that k+1 <= d <= k+m-1. Larger the d, the better the savings.
Type:Integer
Required:No.
Default:k+m-1

scalar_mds={jerasure|isa|shec}

Description:scalar_mds specifies the plugin that is used as a building block in the layered construction. It can be one of jerasure, isa, shec
Type:String
Required:No.
Default:jerasure

technique={technique}

Description:technique specifies the technique that will be picked within the ‘scalar_mds’ plugin specified. Supported techniques are ‘reed_sol_van’, ‘reed_sol_r6_op’, ‘cauchy_orig’, ‘cauchy_good’, ‘liber8tion’ for jerasure, ‘reed_sol_van’, ‘cauchy’ for isa and ‘single’, ‘multiple’ for shec.
Type:String
Required:No.
Default:reed_sol_van (for jerasure, isa), single (for shec)

crush-root={root}

Description:The name of the crush bucket used for the first step of the CRUSH rule. For intance step take default.
Type:String
Required:No.
Default:default

crush-failure-domain={bucket-type}

Description:Ensure that no two chunks are in a bucket with the same failure domain. For instance, if the failure domain is host no two chunks will be stored on the same host. It is used to create a CRUSH rule step such as step chooseleaf host.
Type:String
Required:No.
Default:host

crush-device-class={device-class}

Description:Restrict placement to devices of a specific class (e.g., ssd or hdd), using the crush device class names in the CRUSH map.
Type:String
Required:No.
Default:

directory={directory}

Description:Set the directory name from which the erasure code plugin is loaded.
Type:String
Required:No.
Default:/usr/lib/ceph/erasure-code

--force

Description:Override an existing profile by the same name.
Type:String
Required:No.

Notion of sub-chunks

The Clay code is able to save in terms of disk IO, network bandwidth as it is a vector code and it is able t view and manipulate data within a chunk at a finer granularity termed as a sub-chunk. The number of sub-chunks within a chunk for a Clay code is given by:

sub-chunk count = q(k+m)/q, where q=d-k+1

During repair of an OSD, the helper information requested from an available OSD is only a fraction of a chunk. In fact, the number of sub-chunks within a chunk that are accessed during repair is given by:

repair sub-chunk count = sub-chunk count / q

Examples

  1. For a configuration with k=4, m=2, d=5, the sub-chunk count is 8 and the repair sub-chunk count is 4. Therefore, only half of a chunk is read during repair.
  2. When k=8, m=4, d=11 the sub-chunk count is 64 and repair sub-chunk count is 16. A quarter of a chunk is read from an available OSD for repair of a failed chunk.

How to choose a configuration given a workload

Only a few sub-chunks are read of all the sub-chunks within a chunk. These sub-chunks are not necessarily stored consecutively within a chunk. For best disk IO performance, it is helpful to read contiguous data. For this reaspn, it is suggested that you choose stripe-size such that the sub-chunk size is sufficiently large.

For a given stripe-size (that’s fixed based on a workload), choose k, m, d such that:

sub-chunk size = stripe-size / (k*sub-chunk count) = 4KB, 8KB, 12KB ...
  1. For large size workloads for which the stripe size is large, it is easy to choose k, m, d. For example consider a stripe-size of size 64MB, choosing k=16, m=4 and d=19 will result in a sub-chunk count of 1024 and a sub-chunk size of 4KB.
  2. For small size workloads k=4, m=2 is a good configuration that provides both network and disk IO benefits.

Comparisons with LRC

Locally Recoverable Codes (LRC) are also designed in order to save in terms of network bandwidth, disk IO during single OSD recovery. However, the focus in LRCs is to keep the number of OSDs contacted during repair (d) to be minimal, but this comes at the cost of storage overhead. The clay code has a storage overhead m/k. In the case of an lrc, it stores (k+m)/d parities in addition to the m parities resulting in a storage overhead (m+(k+m)/d)/k. Both clay and lrc can recover from the failure of any m OSDs.

Parameters disk IO, storage overhead (LRC) disk IO, storage overhead (CLAY)
(k=10, m=4) 7 * S, 0.6 (d=7) 3.25 * S, 0.4 (d=13)
(k=16, m=4) 4 * S, 0.5 (d=5) 4.75 * S, 0.25 (d=19)

where S is the amount of data stored of single OSD being recovered.