Browse Prior Art Database

Reducing Impact of Checkpointing in Publish/Subscribe System

IP.com Disclosure Number: IPCOM000123671D
Original Publication Date: 1999-Mar-01
Included in the Prior Art Database: 2005-Apr-05
Document File: 1 page(s) / 38K

Publishing Venue

IBM

Related People

Bolam, S: AUTHOR [+5]

Abstract

Checkpointing persistent state may be achived using a storage medium such as an MQSeries* queue which provides both serial and direct access. The MQSeries Publish/Subscribe function uses this method to make persisent the subscriptions known by each broker. The act of checkpointing this state can impair system performance since generally a broker will need to stop processing new publications while the state is being checkpointed. The MQSeries Publish/Subscribe function lessens the impact of checkpointing by dividing the subscription state into a number of segments. Each segment of the state can be checkpointed individually and generally each segment will require checkpointing at different times from the other segments.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 86% of the total text.

Reducing Impact of Checkpointing in Publish/Subscribe System

   Checkpointing persistent state may be achived using a
storage medium such as an MQSeries* queue which provides both serial
and direct access.  The MQSeries Publish/Subscribe function uses this
method to make persisent the subscriptions known by each broker.  The
act of checkpointing this state can impair system performance since
generally a broker will need to stop processing new publications
while the state is being checkpointed.  The MQSeries
Publish/Subscribe function lessens the impact of checkpointing by
dividing the subscription state into a number of segments.  Each
segment of the state can be checkpointed individually and generally
each segment will require checkpointing at different times from the
other segments.  This has the effect of spreading the load of
checkpointing so that its effect on system response times is less
noticeable.

   Nevertheless brokers may have many thousands of
subscriptions lodged at them.  For example, if each subscription
consumed 256 bytes of queue space and a broker had 50000
subscriptions then, during checkpoint 10 MB of data would need to be
rewritten to the storage medium within a single unit of work.  This
imposes an unreasonable burden on the resources of the underlying
queue manager.

   This problem can be significantly reduced by using a
hashing function to evenly distribute each subscription to one of N
segments each of which can be checkpointed independ...