Data processing: database and file management or data structures – Database design – Data structure types
Reexamination Certificate
2001-06-27
2003-05-13
Breene, John (Department: 2177)
Data processing: database and file management or data structures
Database design
Data structure types
C707S793000, C707S793000, C711S119000
Reexamination Certificate
active
06564230
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to techniques for reducing the penalty associated with one node requesting data from a data store when the most recent version of the requested data resides in the cache of another node.
BACKGROUND OF THE INVENTION
To improve scalability, some database systems permit more than one database server (each running separately) to concurrently access shared storage such as stored on disk media. Each database server has a cache for caching shared resources, such as disk blocks. Such systems are referred to herein as parallel server systems.
One problem associated with parallel server systems is the potential for what are referred to as “pings”. A ping occurs when the version of a resource that resides in the cache of one server must be supplied to the cache of a different server. Thus, a ping occurs when, after a database server A modifies resource x in its cache, a database server B requires resource x for modification. Database servers A and B would typically run on different nodes, but in some cases might run on the same node.
One approach to handling pings is referred to herein as the “disk intervention” approach. The disk intervention approach uses a disk as intermediary storage to transfer the latest version of the resource between two caches. Thus, in the example given above, the disk intervention approach requires database server 1 to write its cache version of resource x to disk, and for database server 2 to retrieve this version from disk into its cache. The disk intervention approach's reliance on two disk I/Os per inter-server transfer of a resource limits the scalability of parallel server systems. Specifically, the disk I/Os required to handle a ping are relatively expensive and time consuming, and the more database servers that are added to the system, the higher the number of pings.
However, the disk intervention approach does provide for relatively efficient recovery from single database server failures, in that such recovery only needs to apply the recovery (redo) log of the failed database server. Applying the redo log of the failed database server ensures that all of the committed changes that transactions on the failed database server made to the resources in the cache of the failed server are recovered. The use of redo logs during recovery are described in detail in U.S. patent application Ser. No. 08/784,611, entitled “CACHING DATA IN RECOVERABLE OBJECTS”, filed on Jan. 21, 1997, the contents of which are incorporated herein by reference.
Parallel server systems that employ the disk intervention approach typically use a protocol in which all global arbitration regarding resource access and modifications is performed by a Distributed Lock Manager (DLM). The operation of an exemplary DLM is described in detail in U.S. patent application Ser. No. 08/669,689, entitled “METHOD AND APPARATUS FOR LOCK CACHING”, filed on Jun. 24, 1996, the contents of which are incorporated herein by reference.
In typical Distributed Lock Manager systems, information pertaining to any given resource is stored in a lock object that corresponds to the resource. Each lock object is stored in the memory of a single node. The lock manager that resides on the node on which a lock object is stored is referred to as the Master of that lock object and the resource it covers.
In systems that employ the disk intervention approach to handling pings, pings tend to involve the DLM in a variety of lock-related communications. Specifically, when a database server (the “requesting server”) needs to access a resource, the database server checks to see whether it has the desired resource locked in the appropriate mode: either shared in case of a read, or exclusive in case of a write. If the requesting database server does not have the desired resource locked in the right mode, or does not have any lock on the resource, then the requesting server sends a request to the Master for the resource to acquire the lock in specified mode.
The request made by the requesting database server may conflict with the current state of the resource (e.g. there could be another database server which currently holds an exclusive lock on the resource). If there is no conflict, the Master for the resource grants the lock and registers the grant. In case of a conflict, the Master of the resource initiates a conflict resolution protocol. The Master of the resource instructs the database server that holds the conflicting lock (the “Holder”) to downgrade its lock to a lower compatible mode.
Unfortunately, if the Holder (e.g. database server A) currently has an updated (“dirty”) version of the desired resource in its cache, it cannot immediately downgrade its lock. In order to downgrade its lock, database server A goes through what is referred to as a “hard ping” protocol. According to the hard ping protocol, database server A forces the redo log associated with the update to be written to disk, writes the resource to disk, downgrades its lock and notifies the Master that database server A is done. Upon receiving the notification, the Master registers the lock grant and notifies the requesting server that the requested lock has been granted. At this point, the requesting server B reads the resource into its cache from disk.
As described above, the disk intervention approach does not allow a resource that has been updated by one database server (a “dirty resource”) to be directly shipped to another database server. Such direct shipment is rendered unfeasible due to recovery related problems. For example, assume that a resource is modified at database server A, and then is shipped directly to database server B. At database server B, the resource is also modified and then shipped back to database server A. At database server A, the resource is modified a third time. Assume also that each server stores all redo logs to disk before sending the resource to another server to allow the recipient to depend on prior changes.
After the third update, assume that database server A dies. The log of database server A contains records of modifications to the resource with a hole. Specifically, server A's log does not include those modifications which were done by database server B. Rather, the modifications made by server B are stored in the database server B's log. At this point, to recover the resource, the two logs must be merged before being applied. This log merge operation, if implemented, would require time and resources proportional to the total number of database servers, including those that did not fail.
The disk intervention approach mentioned above avoids the problem associated with merging recovery logs after a failure, but penalizes the performance of steady state parallel server systems in favor of simple and efficient recovery. The direct shipment approach avoids the overhead associated with the disk intervention approach, but involves complex and nonscalable recovery operations in case of failures.
Based on the foregoing, it is clearly desirable to provide a system and method for reducing the overhead associated with a ping without severely increasing the complexity or duration of recovery operations.
SUMMARY OF THE INVENTION
A method and apparatus are provided for transferring a resource from the cache of one database server to the cache of another database server without first writing the resource to disk. When a database server (Requestor) desires to modify a resource, the Requestor asks for the current version of the resource. The database server that has the current version (Holder) directly ships the current version to the Requestor. Upon shipping the version, the Holder loses permission to modify the resource, but continues to retain a copy of the resource in memory. When the retained version of the resource, or a later version thereof, is written to disk, the Holder can discard the retained version of the resource. Otherwise, the Holder does not discard the retained version. In the case of a server failure, the prior copies of all resources with modific
Bamford Roger J.
Klots Boris
Breene John
Hickman Palermo & Truong & Becker LLP
Holmes Craig G.
Lewis Cheryl
Oracle Corporation
LandOfFree
Transferring a resource between caches of different nodes does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Transferring a resource between caches of different nodes, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Transferring a resource between caches of different nodes will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3070830