Browse Prior Art Database

Partial block read coherency mechanism in a distributed database model

IP.com Disclosure Number: IPCOM000199642D
Publication Date: 2010-Sep-13
Document File: 3 page(s) / 18K

Publishing Venue

The IP.com Prior Art Database

Abstract

Partial block read coherency mechanism in a distributed database model

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 45% of the total text.

Page 1 of 3

Disclosed is aprocess providing a capability for a partial block read coherency mechanism in a distributed database model. The disclosed process enables a member to proceed without requiring a block from a coherency appliance while portions of the block may be valid in a member's buffer cache. The disclosed process further enables the coherency application to return data from a subset of the block, when a member requires data in a block, which is partially invalidated, thereby reducing network traffic. When reading a full block, only sectors, which are invalid in the member, are returned. Data is copied into place in the existing buffer, resulting in a full valid buffer. The disclosed process also provides a capability for a member to specifically request a portion of a block from the coherency appliance, enabling the member to optimally operate on a portion of the data in the block.

In some distributed database models a coherency agent acts as a directory and cache for the individual database members to provide database consistency. Whenever a member needs to read a data block into its buffer cache, the member registers with the coherency agent and receives from the coherency agent a most recent copy of the data block, when the most recent content of the data block is cached in the coherency agent. When a data block is updated, the changed data block is communicated to the coherency agent. The coherency agent invalidates local copies of the block in caches of other members and retains the latest version of the data block. The coherency agent also provides a centralized locking mechanism.

Typically, the coherency agent is connected to the database members using a commodity- based network infrastructure including Ethernet and/or InfiniBand (IB)1 connections. However, these connections do not have infinite performance and may become limiting factors as the capacity of members and number of members increase. Current generations of IB hardware may constrain scalability of modest clusters.

An upper limit of reads or writes per second saturates an adapter. The underlying protocol of a read or write function is composed of a variety of messages, including small control messages and acknowledgements, as well as larger data messages. The referenced test does not include explicit locking function.

The performance problem scales linearly with increasing page or block size. Some workloads may obtain improved performance from larger page sizes. However, the larger page size would likely reduce scalability of other solutions. Clearly, scale-out solutions may be employed. Support of multiple network adapters within a coherency agent, or multiple coherency agents are potential solutions. However, each coherency agent adds significantly to the cost and complexity of the solution.

The disclosed process typically improves the efficiency of the existing mechanisms.

To improve scalability, the amount of data passed through the network as well as t...