RepoMirrors/ceph
1
0
Fork 0
mirror of https://github.com/ceph/ceph synced 2024-12-22 19:34:30 +00:00
Code Issues Packages Projects Releases Wiki Activity
ff9beb2cf5
ceph/doc/dev/osd_internals/erasure_coding/developer_notes.rst
Loic Dachary 14c31ddf10 doc : erasure code developer notes updates 
* unify conventions to match those used by jerasure ( data chunk = K,
  coding chunk = M, use coding instead of parity, use erasures instead
  of erased )

* make lines 80 characters long

* modify the descriptions to take into account that the chunk rank
  will encoded in the pool name and not on a per object basis

* remove the doxygen link to ErasureCodeInterface because it fails
  doc: asphyxiate does not support class
  http://tracker.ceph.com/issues/6115

* only systematic codes are considered at this point ( all jerasure
  techniques are systematic). Although the API could be extended to
  include non systematic codes, it is probably a case of over
  engineering at this point.

* add link to
  http://tracker.ceph.com/issues/6113
  add ceph osd pool create [name] [key=value]

* update the plugin system description to match the proposed
  implementation http://tracker.ceph.com/issues/5877

http://tracker.ceph.com/issues/4929 refs #4929

Reviewed-by: Joao Eduardo Luis <joao.luis@inktank.com>

Signed-off-by: Loic Dachary <loic@dachary.org>
2013-08-27 14:13:56 +02:00

634 lines
25 KiB
ReStructuredText
Raw Blame History

============================
Erasure Code developer notes
============================
Introduction
------------
Each chapter of this document explains an aspect of the implementation
of the erasure code within Ceph. It is mostly based on examples being
explained to demonstrate how things work. It is written as if the
implementation is complete although it may not be the case. For
instance the plugin system and the jerasure plugin are implemented but
the erasure code pool is not.
Reading and writing encoded chunks from and to OSDs
---------------------------------------------------
An erasure coded pool stores each object as K+M chunks. It is divided
into K data chunks and M coding chunks. The pool is configured to have
a size of K+M so that each chunk is stored in an OSD in the acting
set. The rank of the chunks is stored as `an attribute of the pool
<http://tracker.ceph.com/issues/5862>`_ containing the object.
For instance an erasure coded pool is created to use five OSDs ( K+M =
5 ) and sustain the loss of two of them ( M = 2 ).
When the object *NYAN* containing *ABCDEFGHI* is written to it, the
erasure encoding function splits the content in three data chunks,
simply by dividing the content in three : the first contains *ABC*,
the second *DEF* and the last *GHI*. The content will be padded if the
content length is not a multiple of K. The function also creates two
coding chunks : the fourth with *YXY* and the fifth with *GQC*. Each
chunk is stored in an OSD in the acting set. The chunks are stored in
objects that have the same name ( *NYAN* ) but reside on different
OSDs. The order in which the chunks were created must be preserved and
is stored as an attribute of the pool containing the object. Chunk
*1* contains *ABC* and is stored on *OSD5* while chunk *4* contains
*XYY* and is stored on *OSD3*.
::
+-------------------+
name | NYAN |
+-------------------+
content | ABCDEFGHI |
+--------+----------+
|
|
v
+------+------+
+---------------+ encode(3,2) +-----------+
| +--+--+---+---+ |
| | | | |
| +-------+ | +-----+ |
| | | | |
+--v---+ +--v---+ +--v---+ +--v---+ +--v---+
name | NYAN | | NYAN | | NYAN | | NYAN | | NYAN |
+------+ +------+ +------+ +------+ +------+
pool shard | 1 | | 2 | | 3 | | 4 | | 5 |
+------+ +------+ +------+ +------+ +------+
content | ABC | | DEF | | GHI | | YXY | | QGC |
+--+---+ +--+---+ +--+---+ +--+---+ +--+---+
| | | | |
| | | | |
| | +--+---+ | |
| | | OSD1 | | |
| | +------+ | |
| | +------+ | |
| +------>| OSD2 | | |
| +------+ | |
| +------+ | |
| | OSD3 |<----+ |
| +------+ |
| +------+ |
| | OSD4 |<--------------+
| +------+
| +------+
+----------------->| OSD5 |
+------+
When the object *NYAN* is read from the erasure coded pool, the
decoding function reads three chunks : chunk *1* containing *ABC*,
chunk *3* containing *GHI* and chunk *4* containing *YXY* and rebuild
the original content of the object *ABCDEFGHI*. The decoding function
is informed that the chunks *2* and *5* are missing. The chunk *5*
could not be read because the *OSD4* is *out*. The decoding function
is called as soon as three chunks are read : *OSD2* was the slowest
and its chunk was not taken into account.
::
+-------------------+
name | NYAN |
+-------------------+
content | ABCDEFGHI |
+--------+----------+
^
|
|
+------+------+
| decode(3,2) |
| erasures 2,5|
+-------------->| |
| +-------------+
| ^ ^
| | +-----+
| | |
+--+---+ +------+ +--+---+ +--+---+
name | NYAN | | NYAN | | NYAN | | NYAN |
+------+ +------+ +------+ +------+
pool shard | 1 | | 2 | | 3 | | 4 |
+------+ +------+ +------+ +------+
content | ABC | | DEF | | GHI | | YXY |
+--+---+ +--+---+ +--+---+ +--+---+
^ ^ ^ ^
| | | |
| | +--+---+ |
| | | OSD1 | |
| | +------+ |
| | +------+ |
| SLOW +-------| OSD2 | |
| +------+ |
| +------+ |
| | OSD3 |-----+
| +------+
| +------+
| | OSD4 | OUT
| +------+
| +------+
+------------------| OSD5 |
+------+
Interrupted full writes
-----------------------
In an erasure coded pool the primary OSD in the up set receives all
write operations. It is responsible for encoding the payload into K+M
chunks and send them to the OSDs in the up set. It is also responsible
for maintaining an authoritative version of the placement group logs.
::
primary
+---OSD 1---+
| log |
| |
|+----+ |
||D1v1| 1,1 |
|+----+ |
+-----------+
+---OSD 2---+
|+----+ log |
||D2v1| 1,1 |
|+----+ |
+-----------+
+---OSD 3---+
| log |
| |
|+----+ |
||C1v1| 1,1 |
|+----+ |
+-----------+
An erasure coded placement group has been created with K = 2 + M = 1
and is supported by three OSDs, two for K and one for M. The acting
set of the placement group is made of *OSD 1*, *OSD 2* and *OSD 3*. An
object has been encoded and stored in the OSDs : the chunk D1v1
(i.e. Data chunk number 1 version 1) is on *OSD 1*, D2v1 on *OSD 2*
and C1v1 (i.e. Coding chunk number 1 version 1) on *OSD 3*. The
placement group logs on each OSD are in sync at epoch 1 version 1
(i.e. 1,1).
::
primary
+---OSD 1---+
|+----+ log |
||D1v2| 1,2 |<----------------- WRITE FULL
|+----+ |
|+----+ |
||D1v1| 1,1 |
|+----+ |
+++---------+
|| +---OSD 2---+
|| +----+ |+----+ log |
|+-->D2v2| ||D2v1| 1,1 |
| +----+ |+----+ |
| +-----------+
| +---OSD 3---+
| |+----+ log |
+---------->|C1v2| 1,2 |
|+----+ |
|+----+ |
||C1v1| 1,1 |
|+----+ |
+-----------+
*OSD 1* is the primary and receives a WRITE FULL from a client, which
means the payload is to replace the object entirely instead of only
overwriting a portion of it. Version two of the object is created
to override version one. *OSD 1* encodes the payload into three
chunks : D1v2 (i.e. Data chunk number 1 version 2) will be on *OSD 1*,
D2v2 on *OSD 2* and C1v2 (i.e. Coding chunk number 1 version 2) on
*OSD 3*. Each chunk is sent to the target OSD, including the primary
OSD which is responsible for storing chunks in addition to handling
write operations and maintaining an authoritative version of the
placement group logs. When an OSD receives the message instructing it
to write the chunk, it also creates a new entry in the placement group
logs to reflect the change. For instance, as soon as *OSD 3* stores
*C1v2*, it adds the entry 1,2 ( i.e. epoch 1, version 2 ) to its
logs. Because the OSDs work asynchronously, some chunks may still be
in flight ( such as *D2v2* ) while others are acknowledged and on disk
( such as *C1v1* and *D1v1* ). ::
primary
+---OSD 1---+
|+----+ log |
||D1v2| 1,2 |<----------------- WRITE FULL
|+----+ |
|+----+ |
||D1v1| 1,1 |
|+----+ |
+++---------+
|| +---OSD 2---+
|| |+----+ log |
|+--------->|D2v2| 1,2 |
| |+----+ |
| |+----+ |
| ||D2v1| 1,1 |
| |+----+ |
| +-----------+
| +---OSD 3---+
| |+----+ log |
+---------->|C1v2| 1,2 |
|+----+ |
|+----+ |
||C1v1| 1,1 |
|+----+ |
+-----------+
If all goes well, the chunks are acknowledged on each OSD in the
acting set and the logs' *last_complete* pointer can move from
*1,1* to *1,2* and the files used to store the chunks of the previous
version of the object can be removed : *D1v1* on *OSD 1*, *D2v1* on
*OSD 2* and *C1v1* on *OSD 3*.
::
+---OSD 1---+
| |
| DOWN |
| |
+-----------+
+---OSD 2---+
|+----+ log |
||D2v1| 1,1 |
|+----+ |
+-----------+
+---OSD 3---+
|+----+ log |
||C1v2| 1,2 |
|+----+ |
|+----+ |
||C1V1| 1,1 |
|+----+ |
primary +-----------+
+---OSD 4---+
| log |
| 1,1 |
| |
+-----------+
But accidents happen. If *OSD 1* goes down while *D2v2* is still in
flight, the object's version 2 is partially written : *OSD 3* has
one chunk but does not have enough to recover. It lost two chunks :
*D1v2* and *D2v2* but the erasure coding parameters K = 2 + M = 1
requires that at least two chunks are available to rebuild the
third. *OSD 4* becomes the new primary and finds that the
*last_complete* log entry ( i.e. all objects before this entry were
known to be available on all OSDs in the previous acting set ) is
*1,1* and will be the head of the new authoritative log.
::
+---OSD 2---+
|+----+ log |
||D2v1| 1,1 |
|+----+ |
+-----------+
+---OSD 3---+
|+----+ log |
||C1V1| 1,1 |
|+----+ |
primary +-----------+
+---OSD 4---+
| log |
| 1,1 |
| |
+-----------+
The log entry *1,2* found on *OSD 3* is divergent from the new
authoritative log provided by *OSD 4* : it is discarded and the file
containing the *C1v2* chunk is removed.
::
+---OSD 2---+
|+----+ log |
||D2v1| 1,1 |
|+----+ |
+-----------+
+---OSD 3---+
|+----+ log |
||C1V1| 1,1 |
|+----+ |
primary +-----------+
+---OSD 4---+
|+----+ log |
||D1v1| 1,1 |
|+----+ |
+-----------+
The *D1v1* chunk is rebuilt with the *decode* function of the erasure
coding library during scrubbing and stored on the new primary *OSD 4*.
Interrupted append
------------------
An object is coded in stripes, either because they are too big or
because they are created with multiple operations instead of a single
full write. A single stripe will exist/exists in the case of a full
write, assuming the object size is not too large to encode in memory.
When appending to an existing object, the stripe size is retrieved
from the attributes of the object. It applies, for instance, when
*rgw* writes an object with sequence of append instead of a single
write. ::
primary
+---OSD 1---+
|+-s1-+ log |
||S1D1| 1,2 |<----------------- APPEND
||----| |
||S2D1| 1,1 |
|+----+ |
+++---------+
|| +---OSD 2---+
|| +-s2-+ |+-s2-+ log |
|+-->S2D2| ||S1D2| 1,1 |
| +----+ |+----+ |
| +-----------+
| +---OSD 3---+
| |+-s3-+ log |
+---------->|S1C1| 1,2 |
||----| |
||S2C1| 1,1 |
|+----+ |
+-----------+
*OSD 1* is the primary and receives an APPEND from a client, meaning
the payload is to be appended at the end of the object. *OSD 1*
encodes the payload into three chunks : S2D1 (i.e. Stripe two data
chunk number 1 ) will be in s1 ( shard 1 ) on *OSD 1*, S2D2 in s2 on
*OSD 2* and S2C1 (i.e. Stripe two coding chunk number 1 ) in s3 on
*OSD 3*. Each chunk is sent to the target OSD, including the primary
OSD which is responsible for storing chunks in addition to handling
write operations and maintaining an authoritative version of the
placement group logs. When an OSD receives the message instructing it
to write the chunk, it also creates a new entry in the placement group
logs to reflect the change. For instance, as soon as *OSD 3* stores
*S2C1*, it adds the entry 1,2 ( i.e. epoch 1, version 2 ) to its
logs. The log entry also carries the nature of the operation: in this
case 1,2 is an APPEND where 1,1 was a CREATE. Because the OSDs work
asynchronously, some chunks may still be in flight ( such as *S2D2* )
while others are acknowledged and on disk ( such as *S2D1* and *S2C1*
).
::
+---OSD 1---+
| |
| DOWN |
| |
+-----------+
+---OSD 2---+
|+-s2-+ log |
||S1D2| 1,1 |
|+----+ |
+-----------+
+---OSD 3---+
|+-s3-+ log |
||S1C1| 1,2 |
||----| |
||S2C1| 1,1 |
|+----+ |
primary +-----------+
+---OSD 4---+
| log |
| 1,1 |
| |
+-----------+
If *OSD 1* goes down while *S2D2* is still in flight, the payload is
partially appended : s3 ( shard 3) in *OSD 3* has one chunk but does
not have enough to recover because s1 and s2 don't have it. Two chunks
were lost (*S2D1* and S2D2) but the erasure coding parameters K = 2 +
M = 1 requires that at least two chunks are available to rebuild the
third. *OSD 4* becomes the new primary and finds that the
*last_complete* log entry ( i.e. all objects before this entry were
known to be available on all OSDs in the previous acting set ) is
*1,1* and will be the head of the new authoritative log. ::
+---OSD 2---+
|+-s2-+ log |
||S1D2| 1,1 |
|+----+ |
+-----------+
+---OSD 3---+
|+-s3-+ log |
||S1C1| 1,1 |
|+----+ |
primary +-----------+
+---OSD 4---+
| log |
| 1,1 |
| |
+-----------+
The log entry *1,2* found on *OSD 3* is divergent from the new
authoritative log provided by *OSD 4* : it is discarded and the file
containing the *S2C1* chunk is truncated to the nearest multiple of
the stripe size.
Erasure code library
--------------------
See also `the corresponding tracker issue <http://tracker.ceph.com/issues/5877>`_
Using `Reed-Solomon <https://en.wikipedia.org/wiki/Reed_Solomon>`_,
with parameters K+M, object O is encoded by dividing it into chunks O1,
O2, ... OM and computing coding chunks P1, P2, ... PK. Any K chunks
out of the available K+M chunks can be used to obtain the original
object. If data chunk O2 or coding chunk P2 are lost, they can be
repaired using any K chunks out of the K+M chunks. If more than M
chunks are lost, it is not possible to recover the object.
Reading the original content of object O could be a simple
concatenation of O1, O2, ... OM, because the plugins are using
`systematic codes
<http://en.wikipedia.org/wiki/Systematic_code>`_. Otherwise the chunks
must be given to the erasure code library to retrieve the content of
the object.
Reed-Solomon is significantly more expensive to encode than fountain
codes with the current `jerasure implementation
<http://web.eecs.utk.edu/~plank/plank/papers/CS-08-627.html>`_. However
`gf-complete
<http://web.eecs.utk.edu/~plank/plank/papers/CS-13-703.html>`_ that
will be used in the upcoming version of jerasure is twice faster and
the difference becomes negligible. The difference is even more
important when an object is divided in hundreds or more chunks, but
Ceph will typically be used with less than 32 chunks.
Performance depend on the parameters to the encoding functions and
is also influenced by the packet sizes used when calling the encoding
functions ( for Cauchy or Liberation for instance ): smaller packets
means more calls and more overhead.
Although Reed-Solomon is provided as a default, Ceph uses it via an
`abstract API <http://tracker.ceph.com/issues/5878>`_ designed to
allow each pool to choose the plugin that implements it using
`key=value pairs when creating the pool
<http://tracker.ceph.com/issues/6113>`_.
::
ceph osd pool create <pool> \
erasure-code-directory=<dir> \
erasure-code-plugin=<plugin>
The *<plugin>* is dynamically loaded from *<dir>* (defaults to
*/usr/lib/ceph/erasure-code* ) and expected to implement the
*int __erasure_code_init(char *plugin_name)* function
which is responsible for registering an object derived from
*ErasureCodePlugin* in the registry :
::
ErasureCodePluginRegistry::add(plugin_name,
new ErasureCodePluginExample());
The *ErasureCodePlugin* derived object must provide a factory method
from which the concrete implementation of the *ErasureCodeInterface*
object can be generated:
::
virtual int factory(const map<std::string,std::string> &parameters,
ErasureCodeInterfaceRef *erasure_code) {
*erasure_code = ErasureCodeInterfaceRef(new ErasureCodeExample(parameters));
return 0;
}
The *parameters* is the list of *key=value* pairs that were set when the pool
was created. Each *key* must be prefixed with erasure-code to avoid name collisions
::
ceph osd pool create <pool> \
erasure-code-directory=<dir> \ # mandatory
erasure-code-plugin=jerasure \ # mandatory
erasure-code-m=10 \ # optional and plugin dependant
erasure-code-k=3 \ # optional and plugin dependant
erasure-code-technique=reed_sol_van \ # optional and plugin dependant
Erasure code jerasure plugin
----------------------------
The parameters interpreted by the jerasure plugin are:
::
ceph osd pool create <pool> \
erasure-code-directory=<dir> \ # plugin directory absolute path
erasure-code-plugin=jerasure \ # plugin name (only jerasure)
erasure-code-k=<k> \ # data chunks (default 2)
erasure-code-m=<m> \ # coding chunks (default 2)
erasure-code-technique=<technique> \ # coding technique
The coding techniques can be chosen among *reed_sol_van*,
*reed_sol_r6_op*, *cauchy_orig*, *cauchy_good*, *liberation*,
*blaum_roth* and *liber8tion*.
Scrubbing
---------
See also `Refactor scrub to use PGBackend methods <http://tracker.ceph.com/issues/5861>`_
The simplest form of scrubbing is to check with each OSDs holding a
chunk if it exists locally. If more thank M chunks are missing the
object is marked as lost. If up to M chunks are missing they are
repaired and written to the relevant OSDs.
From time to time it may make sense to attempt to read an object,
using all of its chunks. If the decode function fails, the object is
lost.
Bit flips happen. Not often, but it is possible. Here is `an article
from 2011 <http://www.linux-mag.com/id/8794/>`_ also search for "bit
rot" and "bit error rate". To detect corrupted chunks, a checksum
(CRC23C for instance) must be added as an attribute of the file
containing the chunk ( or shard ) so that deep scrubbing can check
that the chunk is valid by recomputing the content of the chunk and
compare it with the signature. BTRFS and ZFS have a CRC32C check
built-in on a per block basis.
Notes
-----
This document is a description of how erasure coding could be
implemented, it does not reflect the current state of the code
base. Possible optimizations are mentionned where relevant but the
first implementation should not include any of them: they are
presented to show that there is a path toward optimization starting
from simple minded implementation.
If the objects are large, it may be impractical to encode and decode
them in memory. However, when using *RBD* a 1TB device is divided in
many individual 4MB objects and *RGW* does the same.
Encoding and decoding is implemented in the OSD. Although it could be
implemented client side for read write, the OSD must be able to encode
and decode on its own when scrubbing.
If a partial read is required, an optimization could be to only fetch
the chunk that contains the data instead of always fetching all
chunks. For instance if *H* is required in the example above, chunk 3
is read if available. Reading 3 chunks is a fallback in case chunk 3 is
not available.
Partial reads and writes
------------------------
If an object is large, reading or writing all of it when changing only
a few bytes is expensive. It is more efficient to only read or write a
subset of the object. When a client writes on an existing object, it
can provide the offset and the length of the write as well as the
payload with the `CEPH_OSD_OP_WRITE
<https://github.com/ceph/ceph/blob/962b64a83037ff79855c5261325de0cd1541f582/src/osd/ReplicatedPG.cc#L2542>`_
operation. It is refered to as *partial write* and is different from
the `CEPH_OSD_OP_WRITEFULL operation
<https://github.com/ceph/ceph/blob/962b64a83037ff79855c5261325de0cd1541f582/src/osd/ReplicatedPG.cc#L2552>`_
which writes the entire object at once.
When using replicas for partial writes or reads, the primary OSD
translates them into read(2) and write(2) POSIX system calls. When
writing, it then forwards the CEPH_OSD_OP_WRITE message to the
replicas and waits for them to acknowledge they are done.
When reading erasure coded objects, at least M chunks must be read and
decoded to extract the desired bytes. If a `systematic code
<https://en.wikipedia.org/wiki/Systematic_code>`_ is used ( i.e. the
data chunks are readable by simple concatenation ) read can be
optimized to use the chunk containing the desired bytes and rely on
the erasure decoding function only if a chunk is missing.
When writing an erasure coded object, changing even one byte requires
that it is encoded again in full.
If Ceph is only used thru the *radosgw* or *librbd*, objects will mostly
have the same size. The *radosgw* user may upload a 1GB object, which will
be divided into smaller 4MB objects behind the scene ( or whatever is
set with *rgw obj stripe size* ). If a KVM is attached a 10GB RBD block
device, it will also be divided into smaller 4BM objects ( or whatever
size is given to the --stripe-unit argument when creating the RBD
block ). In both cases, writing one byte at the beginning will only
require to encode the first object and not all of them.
Objects can be further divided into stripes to reduce the overhead of
partial writes. For instance:
::
+-----------------------+
|+---------------------+|
|| stripe 0 ||
|| [0,N) ||
|+---------------------+|
|+---------------------+|
|| stripe 1 ||
|| [N,N*2) ||
|+---------------------+|
|+---------------------+|
|| stripe 3 [N*2,len) ||
|+---------------------+|
+-----------------------+
object of size len
Each stripe is encoded independantly and the same OSDs are used for
all of them. For instance, if stripe 0 is encoded into 3 chunks on
OSDs 5, 8 and 9, stripe 1 is also encoded into 3 chunks on the same
OSDs. The size of a stripe is stored as an attribute of the object.
When writing one byte at offset N, instead of re-encoding the whole
object it is enough to re-encode the stripe that contains it.
Reference in New Issue View Git Blame Copy Permalink