mirror of https://github.com/ceph/ceph
152 lines
5.6 KiB
ReStructuredText
152 lines
5.6 KiB
ReStructuredText
===============
|
|
Deduplication
|
|
===============
|
|
|
|
|
|
Introduction
|
|
============
|
|
|
|
Applying data deduplication on an existing software stack is not easy
|
|
due to additional metadata management and original data processing
|
|
procedure.
|
|
|
|
In a typical deduplication system, the input source as a data
|
|
object is split into multiple chunks by a chunking algorithm.
|
|
The deduplication system then compares each chunk with
|
|
the existing data chunks, stored in the storage previously.
|
|
To this end, a fingerprint index that stores the hash value
|
|
of each chunk is employed by the deduplication system
|
|
in order to easily find the existing chunks by comparing
|
|
hash value rather than searching all contents that reside in
|
|
the underlying storage.
|
|
|
|
There are many challenges in order to implement deduplication on top
|
|
of Ceph. Among them, two issues are essential for deduplication.
|
|
First is managing scalability of fingerprint index; Second is
|
|
it is complex to ensure compatibility between newly applied
|
|
deduplication metadata and existing metadata.
|
|
|
|
Key Idea
|
|
========
|
|
1. Content hashing (Double hashing): Each client can find an object data
|
|
for an object ID using CRUSH. With CRUSH, a client knows object's location
|
|
in Base tier.
|
|
By hashing object's content at Base tier, a new OID (chunk ID) is generated.
|
|
Chunk tier stores in the new OID that has a partial content of original object.
|
|
|
|
Client 1 -> OID=1 -> HASH(1's content)=K -> OID=K ->
|
|
CRUSH(K) -> chunk's location
|
|
|
|
|
|
2. Self-contained object: The external metadata design
|
|
makes difficult for integration with storage feature support
|
|
since existing storage features cannot recognize the
|
|
additional external data structures. If we can design data
|
|
deduplication system without any external component, the
|
|
original storage features can be reused.
|
|
|
|
More details in https://ieeexplore.ieee.org/document/8416369
|
|
|
|
Design
|
|
======
|
|
|
|
.. ditaa::
|
|
|
|
+-------------+
|
|
| Ceph Client |
|
|
+------+------+
|
|
^
|
|
Tiering is |
|
|
Transparent | Metadata
|
|
to Ceph | +---------------+
|
|
Client Ops | | |
|
|
| +----->+ Base Pool |
|
|
| | | |
|
|
| | +-----+---+-----+
|
|
| | | ^
|
|
v v | | Dedup metadata in Base Pool
|
|
+------+----+--+ | | (Dedup metadata contains chunk offsets
|
|
| Objecter | | | and fingerprints)
|
|
+-----------+--+ | |
|
|
^ | | Data in Chunk Pool
|
|
| v |
|
|
| +-----+---+-----+
|
|
| | |
|
|
+----->| Chunk Pool |
|
|
| |
|
|
+---------------+
|
|
Data
|
|
|
|
|
|
Pool-based object management:
|
|
We define two pools.
|
|
The metadata pool stores metadata objects and the chunk pool stores
|
|
chunk objects. Since these two pools are divided based on
|
|
the purpose and usage, each pool can be managed more
|
|
efficiently according to its different characteristics. Base
|
|
pool and the chunk pool can separately select a redundancy
|
|
scheme between replication and erasure coding depending on
|
|
its usage and each pool can be placed in a different storage
|
|
location depending on the required performance.
|
|
|
|
Regarding how to use, please see ``osd_internals/manifest.rst``
|
|
|
|
Usage Patterns
|
|
==============
|
|
|
|
The different Ceph interface layers present potentially different oportunities
|
|
and costs for deduplication and tiering in general.
|
|
|
|
RadosGW
|
|
-------
|
|
|
|
S3 big data workloads seem like a good opportunity for deduplication. These
|
|
objects tend to be write once, read mostly objects which don't see partial
|
|
overwrites. As such, it makes sense to fingerprint and dedup up front.
|
|
|
|
Unlike cephfs and rbd, radosgw has a system for storing
|
|
explicit metadata in the head object of a logical s3 object for
|
|
locating the remaining pieces. As such, radosgw could use the
|
|
refcounting machinery (``osd_internals/refcount.rst``) directly without
|
|
needing direct support from rados for manifests.
|
|
|
|
RBD/Cephfs
|
|
----------
|
|
|
|
RBD and CephFS both use deterministic naming schemes to partition
|
|
block devices/file data over rados objects. As such, the redirection
|
|
metadata would need to be included as part of rados, presumably
|
|
transparently.
|
|
|
|
Moreover, unlike radosgw, rbd/cephfs rados objects can see overwrites.
|
|
For those objects, we don't really want to perform dedup, and we don't
|
|
want to pay a write latency penalty in the hot path to do so anyway.
|
|
As such, performing tiering and dedup on cold objects in the background
|
|
is likely to be preferred.
|
|
|
|
One important wrinkle, however, is that both rbd and cephfs workloads
|
|
often feature usage of snapshots. This means that the rados manifest
|
|
support needs robust support for snapshots.
|
|
|
|
RADOS Machinery
|
|
===============
|
|
|
|
For more information on rados redirect/chunk/dedup support, see ``osd_internals/manifest.rst``.
|
|
For more information on rados refcount support, see ``osd_internals/refcount.rst``.
|
|
|
|
Status and Future Work
|
|
======================
|
|
|
|
At the moment, there exists some preliminary support for manifest
|
|
objects within the OSD as well as a dedup tool.
|
|
|
|
RadosGW data warehouse workloads probably represent the largest
|
|
opportunity for this feature, so the first priority is probably to add
|
|
direct support for fingerprinting and redirects into the refcount pool
|
|
to radosgw.
|
|
|
|
Aside from radosgw, completing work on manifest object support in the
|
|
OSD particularly as it relates to snapshots would be the next step for
|
|
rbd and cephfs workloads.
|
|
|