Restartable database loads using parallel data streams

Details Restartable database loads using parallel data streams Restartable database loads using parallel data streams

2001-03-28

2004-12-21

Baderman, Scott (Department: 2114)

Error detection/correction and fault detection/recovery

Data processing system error or fault handling

Reliability and availability

C714S012000

Reexamination Certificate

active

06834358

ABSTRACT:

BACKGROUND
Relational Database Management Systems (RDBMS) have become an integral part of enterprise information processing infrastructures throughout the world. An RDBMS
100
, as shown in
FIG. 1
, maintains relational data structures called “relational tables,” or simply “tables”
105
. Tables
105
consist of related data values known as “columns” (or “attributes”) which form “rows” (or “tuples”).
An RDBMS “server”
110
is a hardware and/or software entity responsible for supporting the relational paradigm. As its name implies, the RDBMS server provides services to other programs, i.e., it stores, retrieves, organizes and manages data. A software program that uses the services provided by the RDBMS Server is known as a “client”
115
.
In many cases, an enterprise will store real-time data in an operational data store (ODS)
200
, illustrated in
FIG. 2
, which is designed to efficiently handle a large number of small transactions, such as sales transactions, in a short amount of time. If the enterprises wishes to perform analysis of the data stored in the ODS, it may move the data to a data warehouse
205
, which is designed to handle a relatively small number of very large transactions that require reasonable, but not necessarily instantaneous response times.
To accomplish this, data is “imported,” or “loaded” (block
210
) from various external sources, such as the ODS
200
, into the data warehouse
205
. Once the data is inside the data warehouse
205
, it can be manipulated and queried. Similarly, the data is sometimes “unloaded” or “exported” from the data warehouse
205
into the ODS
200
or into another data store. Since both load and unload processes share many similarities, in terms of the processing they perform, they will be referred to hereinafter as “database loads” or “loads.”
A database load is typically performed by a special purpose program called a “utility.” In most cases the time required to perform a database load is directly proportional to the amount of data being transferred. Consequently, loading or unloading “Very Large Databases” (i.e. databases containing many gigabytes of data) creates an additional problem—increased risk of failure. The longer a given load runs, the higher the probability is that it will be unexpectedly interrupted by a sudden hardware or software failure on either the client
115
or the server
110
. If such a failure occurs, some or all of the data being loaded or unloaded may be lost or unsuitable for use and it may be necessary to restart the load or unload process.
“Parallel Processing,” a computing technique in which computations are performed simultaneously by multiple computing resources, can reduce the amount of time necessary to perform a load by distributing the processing associated with the load across a number of processors. Reducing the load time reduces the probability of failure. Even using parallel processing, however, the amount of data is still very large and errors are still possible.
One traditional approach to handling errors in non-parallel systems is called “mini-batch” or “checkpointing.” Using this approach, the overall processing time for a task is divided into a set of intervals. At the end of each interval, the task enters a “restartable state” called a “checkpoint” and makes a permanent record of this fact. A restartable state is a program state from which processing can be resumed as if it had never been interrupted. If processing is interrupted, it can be resumed from the most recent successful checkpoint without introducing any errors into the final result.
Applying checkpointing to a parallel process is a significant challenge.
SUMMARY
In general, in one aspect, the invention features a method for reducing the restart time for a parallel application. The parallel application includes a plurality of parallel operators. The method includes repeating the following: setting a time interval to a next checkpoint; waiting until the time interval expires; sending checkpoint requests to each of the plurality of parallel operators; and receiving and processing messages from one or more of the plurality of parallel operators.
Implementations of the invention may include one or more of the following. Before entering the repeat loop the method may include receiving a ready message from each of the plurality of parallel operators indicating the parallel operator that originated the message is ready to accept checkpoint requests.
Receiving and processing messages from one or more of the plurality of parallel operators may include receiving a checkpoint information message, including checkpoint information, from one of the plurality of parallel operators and storing the checkpoint information, along with an identifier for the one of the parallel operators, in a checkpoint data store. Receiving and processing messages from one or more of the plurality of parallel operators may include receiving a ready to proceed message from one of the plurality of parallel operators, marking the one of the plurality of parallel operators as ready to proceed, and, if all of the plurality of parallel operators has been marked as ready to proceed, marking a current checkpoint as good. Receiving and processing messages from one or more of the plurality of parallel operators may include receiving a checkpoint reject message from one of the plurality of parallel operators, sending abandon checkpointing messages to the plurality of parallel operators, and scheduling a new checkpoint. Receiving and processing messages from one or more of the plurality of parallel operators may include receiving a recoverable error message from one or more of the plurality of parallel operators, sending abandon checkpointing messages to the plurality of parallel operators, waiting for ready messages from all of the plurality of parallel operators, and scheduling a new checkpoint. Receiving and processing messages from one or more of the plurality of parallel operators may include receiving a non-recoverable error message from one of the plurality of parallel operators, and sending terminate messages to the plurality of parallel operators.
The method may further include restarting the plurality of parallel operators. Restarting may include sending initiate restart messages to the plurality of parallel processors and processing restart messages from the plurality of parallel processors. Processing restart messages may include receiving an information request message from one or more of the plurality of parallel operators, retrieving checkpoint information regarding the one or more of the plurality of parallel operators from the checkpoint data store, and sending the retrieved information to the one of the plurality of parallel operators. Processing restart messages may include receiving a ready to proceed message from one of the plurality of parallel operators, marking the one of the plurality of parallel operators as ready to proceed, and sending proceed messages to all of the plurality of parallel operators if all of the plurality of parallel operators have been marked as ready to proceed. Processing restart messages may comprise receiving an error message from one or more of the plurality of parallel operators and terminating the processing of the plurality of parallel operators.
In general, in another aspect, the invention features a method for one of a plurality of parallel operators to record its state. The method includes receiving a checkpoint request message on a control data stream, waiting to enter a state suitable for checkpointing, and sending a response message on the control data stream.
Implementations of the invention may include one or more of the following. Waiting to enter a state suitable for checkpointing comprises receiving a checkpoint marker on an input data stream, finishing writing data to an output data stream, and sending a checkpoint marker on the output data stream. Waiting to enter a state suitable for checkpointing may comprise waiting for all of the parallel operator's outstanding input/output requests to be

Affiliated with

Also associated with

Rating

Restartable database loads using parallel data streams does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Restartable database loads using parallel data streams, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Restartable database loads using parallel data streams will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-3322481