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

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.