5 CLAY (short for coupled-layer) codes are erasure codes designed to bring about significant savings
6 in terms of network bandwidth and disk IO when a failed node/OSD/rack is being repaired. Let:
8 d = number of OSDs contacted during repair
10 If *jerasure* is configured with *k=8* and *m=4*, losing one OSD requires
11 reading from the *d=8* others to repair. And recovery of say a 1GiB needs
12 a download of 8 X 1GiB = 8GiB of information.
14 However, in the case of the *clay* plugin *d* is configurable within the limits:
18 By default, the clay code plugin picks *d=k+m-1* as it provides the greatest savings in terms
19 of network bandwidth and disk IO. In the case of the *clay* plugin configured with
20 *k=8*, *m=4* and *d=11* when a single OSD fails, d=11 osds are contacted and
21 250MiB is downloaded from each of them, resulting in a total download of 11 X 250MiB = 2.75GiB
22 amount of information. More general parameters are provided below. The benefits are substantial
23 when the repair is carried out for a rack that stores information on the order of
26 +-------------+---------------------------+
27 | plugin | total amount of disk IO |
28 +=============+===========================+
30 +-------------+---------------------------+
31 | clay | d*S/(d-k+1) = (k+m-1)*S/m |
32 +-------------+---------------------------+
34 where *S* is the amount of data stored on a single OSD undergoing repair. In the table above, we have
35 used the largest possible value of *d* as this will result in the smallest amount of data download needed
36 to achieve recovery from an OSD failure.
38 Erasure-code profile examples
39 =============================
41 An example configuration that can be used to observe reduced bandwidth usage::
43 $ ceph osd erasure-code-profile set CLAYprofile \
46 crush-failure-domain=host
47 $ ceph osd pool create claypool 12 12 erasure CLAYprofile
50 Creating a clay profile
51 =======================
53 To create a new clay code profile::
55 ceph osd erasure-code-profile set {name} \
60 [scalar_mds={plugin-name}] \
61 [technique={technique-name}] \
62 [crush-failure-domain={bucket-type}] \
63 [directory={directory}] \
70 :Description: Each object is split into **data-chunks** parts,
71 each of which is stored on a different OSD.
79 :Description: Compute **coding chunks** for each object and store them
80 on different OSDs. The number of coding chunks is also
81 the number of OSDs that can be down without losing data.
89 :Description: Number of OSDs requested to send data during recovery of
90 a single chunk. *d* needs to be chosen such that
91 k+1 <= d <= k+m-1. Larger the *d*, the better the savings.
97 ``scalar_mds={jerasure|isa|shec}``
99 :Description: **scalar_mds** specifies the plugin that is used as a
100 building block in the layered construction. It can be
101 one of *jerasure*, *isa*, *shec*
107 ``technique={technique}``
109 :Description: **technique** specifies the technique that will be picked
110 within the 'scalar_mds' plugin specified. Supported techniques
111 are 'reed_sol_van', 'reed_sol_r6_op', 'cauchy_orig',
112 'cauchy_good', 'liber8tion' for jerasure, 'reed_sol_van',
113 'cauchy' for isa and 'single', 'multiple' for shec.
117 :Default: reed_sol_van (for jerasure, isa), single (for shec)
120 ``crush-root={root}``
122 :Description: The name of the crush bucket used for the first step of
123 the CRUSH rule. For intance **step take default**.
130 ``crush-failure-domain={bucket-type}``
132 :Description: Ensure that no two chunks are in a bucket with the same
133 failure domain. For instance, if the failure domain is
134 **host** no two chunks will be stored on the same
135 host. It is used to create a CRUSH rule step such as **step
142 ``crush-device-class={device-class}``
144 :Description: Restrict placement to devices of a specific class (e.g.,
145 ``ssd`` or ``hdd``), using the crush device class names
152 ``directory={directory}``
154 :Description: Set the **directory** name from which the erasure code
159 :Default: /usr/lib/ceph/erasure-code
163 :Description: Override an existing profile by the same name.
172 The Clay code is able to save in terms of disk IO, network bandwidth as it
173 is a vector code and it is able to view and manipulate data within a chunk
174 at a finer granularity termed as a sub-chunk. The number of sub-chunks within
175 a chunk for a Clay code is given by:
177 sub-chunk count = q\ :sup:`(k+m)/q`, where q=d-k+1
180 During repair of an OSD, the helper information requested
181 from an available OSD is only a fraction of a chunk. In fact, the number
182 of sub-chunks within a chunk that are accessed during repair is given by:
184 repair sub-chunk count = sub-chunk count / q
189 #. For a configuration with *k=4*, *m=2*, *d=5*, the sub-chunk count is
190 8 and the repair sub-chunk count is 4. Therefore, only half of a chunk is read
192 #. When *k=8*, *m=4*, *d=11* the sub-chunk count is 64 and repair sub-chunk count
193 is 16. A quarter of a chunk is read from an available OSD for repair of a failed
198 How to choose a configuration given a workload
199 ==============================================
201 Only a few sub-chunks are read of all the sub-chunks within a chunk. These sub-chunks
202 are not necessarily stored consecutively within a chunk. For best disk IO
203 performance, it is helpful to read contiguous data. For this reason, it is suggested that
204 you choose stripe-size such that the sub-chunk size is sufficiently large.
206 For a given stripe-size (that's fixed based on a workload), choose ``k``, ``m``, ``d`` such that::
208 sub-chunk size = stripe-size / (k*sub-chunk count) = 4KB, 8KB, 12KB ...
210 #. For large size workloads for which the stripe size is large, it is easy to choose k, m, d.
211 For example consider a stripe-size of size 64MB, choosing *k=16*, *m=4* and *d=19* will
212 result in a sub-chunk count of 1024 and a sub-chunk size of 4KB.
213 #. For small size workloads, *k=4*, *m=2* is a good configuration that provides both network
214 and disk IO benefits.
219 Locally Recoverable Codes (LRC) are also designed in order to save in terms of network
220 bandwidth, disk IO during single OSD recovery. However, the focus in LRCs is to keep the
221 number of OSDs contacted during repair (d) to be minimal, but this comes at the cost of storage overhead.
222 The *clay* code has a storage overhead m/k. In the case of an *lrc*, it stores (k+m)/d parities in
223 addition to the ``m`` parities resulting in a storage overhead (m+(k+m)/d)/k. Both *clay* and *lrc*
224 can recover from the failure of any ``m`` OSDs.
226 +-----------------+----------------------------------+----------------------------------+
227 | Parameters | disk IO, storage overhead (LRC) | disk IO, storage overhead (CLAY) |
228 +=================+================+=================+==================================+
229 | (k=10, m=4) | 7 * S, 0.6 (d=7) | 3.25 * S, 0.4 (d=13) |
230 +-----------------+----------------------------------+----------------------------------+
231 | (k=16, m=4) | 4 * S, 0.5 (d=5) | 4.75 * S, 0.25 (d=19) |
232 +-----------------+----------------------------------+----------------------------------+
235 where ``S`` is the amount of data stored of single OSD being recovered.