Keep » History » Version 4

« Previous - Version 4/26 (diff) - Next » - Current version
Anonymous, 04/11/2013 01:02 PM


Keep

Keep is a distributed content-addressable storage system designed for high performance in I/O-bound cluster environments.

Notable design goals and features include:

  • High scalability
  • Node-level redundancy
  • Maximum overall throughput in a busy cluster environment
  • Maximum data bandwidth from client to disk
  • Minimum transaction overhead
  • Elimination of disk thrashing (commonly caused by multiple simultaneous readers)
  • Client-controlled redundancy

Design

The above goals are accomplished by the following design features.

  • Data is transferred directly between the client and the physical node where the disk is installed.
  • Data collections are encoded in large (≤64 MiB) blocks to minimize short read/write operations.
  • Each disk accepts only one block-read/write operation at a time. This prevents disk thrashing and maximizes total throughput when many clients compete for a disk.
  • Storage redundancy is directly controlled, and can be easily verified, by the client simply by reading or writing a block of data on multiple nodes.
  • Data block distribution is computed based on a cryptographic digest of the data block being stored or retrieved. This eliminates the need for a central or synchronized database of block storage locations.

Components

The Keep storage system consists of data block read/write services, SDKs, and management agents.

The responsibilities of the Keep service are:

  • Write data blocks
  • When writing: ensure data integrity by comparing client-supplied cryptographic digest and data
  • Read data blocks (subject to permission, which is determined by the system/metadata DB)
  • Send read/write/error event logs to management agents

The responsibilities of the SDK are:

  • When writing: split data into ≤64 MiB chunks
  • When writing: encode directory trees as manifests
  • When writing: write data to the desired number of nodes to achieve storage redundancy
  • After writing: register a collection with Arvados
  • When reading: parse manifests
  • When reading: verify data integrity by comparing locator to MD5 digest of retrieved data

The responsibilities of management agents are:

  • Verify validity of permission tokens
  • Determine which blocks have higher or lower redundancy than required
  • Monitor disk space and move or delete blocks as needed
  • Collect per-user, per-group, per-node, and per-disk usage statistics

Benefits

Keep offers a variety of major benefits over POSIX file systems and other object file storage systems. This is a summary of some of those benefits:

  • Elimination of Duplication - One of the major storage management problems today is the duplication of data. Often researchers will make copies of data for backup or to re-organize files for different projects. Content addressing eliminates unnecessary duplication because Keep checks when a file is saved to verify if the file already exists. If it does, there is no reason to make a copy.
  • Canonical Records - Content addressing creates clear and verifiable canonical records for files. By combining Keep with the computation system in Arvados, it becomes trivial to verify if you the actual file that was used for a computation. By using a collection to define an entire data set (which could be 100s of terabytes or petabytes), you maintain a permanent and verifiable record of which data were used for each computation. The file that defines a collection is very small relative to the underlying data, so you can make as many as you need.
  • Provenance - The combination of Keep and the computation system make it possible to maintain clear provenance for all the data in the system. This has a number of benefits including making it easy to ascertain how data were derived at any point in time.
  • Easy Management of Temporary Data - As a result of provenance, the system can easily manage temporary data. If you know how a data set or file is was created, you can decide if it is worth it to keep that file in storage. Knowing what pipeline was run on which input data to create gives you that visibility.
  • Flexible Organization - In Arvados, files in Keep can be easily tagged with metadata. The roadmap is that it will be possible to create separate namespaces for metadata for different researchers or research teams. This makes it possible for researchers to organize files in a variety of different ways. Datasets are defined by creating collections. A collection is a text file that stores pointers to other files using the content addresses, and each collection has it's own address. As a result, the same underlying data can be described in a virtually unlimited number of different collections without every re-organizing the physical structure of the data.
  • High-Reliability - By combining content addressing with an object file store, Keep is fault tolerant across drive and even node failures in the system. The Data Manager can control how many replications are maintained in an Arvados cluster depending on the importance of the data and whether or not it has been backed up in another system.
  • Easier Tiering of Storage - The Data Manager in Arvados will be able to manage the distribution of files to storage systems such as a NAS or cloud back up service. Because the files are all content addressed and tracked in the metadata base, if a pipeline is run that asks for a file which is not on the cluster, Arvados can automatically pause that pipeline and move the necessary data on to the cluster for the job. This makes tiering storage transparent to end-users eliminating many of the challenges associated with convincing researchers to re-organize there data.
  • Security and Access Control - Keep can encrypt files on disk and this storage architecture makes the implementation of very fine grain access control significantly easier than traditional POSIX file systems.
  • POSIX Interface - While it's a slower interface, it is possible to mount collections in Keep as POSIX drives in a virtual machine in order to access data in a way that existing tools may expect. Because collections are so flexible, one could easily create many different virtual directory structures for the same underlying files without ever physically moving the files. Over time we think shifting to the APIs and SDK will be a better solution than mounting virtual POSIX drives, but this approach will ease the transition.
  • Data Sharing - Keep makes it much easier to share data between clusters in different data centers or organizations. When files are content addressed the address includes information about the cluster where the file lives. Clusters can be federated so it's possible define collections of data that reside on multiple clusters, which makes it possible to distribute computations across clusters, eliminating the need to physically move the data.