Fast Abstracts Archives . .

A Reliable Real-Time Recovery Manager
for Distributed Real-Time DBMS
in Switching Systems

1. Introduction

Most real-time systems, especially the switching systems[4], keep the shared operational data in the embedded real-time DBMS for real-time processing and reliability. Since the real-time DBMS holds the whole databases in main memory for high performance, preserving the reliable database is an essential requirement. On the system failure, moreover, the database recovery is crucial to the system reliability. In this paper we suggest a real-time recovery method for the main-memory DBMS which reflects the real-time characteristics and overcomes the limitations in the switching systems.

The key idea of the proposed method is to employ a specific manager called Real-Time Recovery Manager(RTRM), which tackles the problems in the systems. In the switching systems, the conventional recovery technique[1,2,3] cannot be adopted due to the I/O overheads on the server processor. The RTRM is devised for pseudo-parallel recovery of the databases and relieves the disk overhead on the server processor.

The RTRM ensures that any erroneous databases can be repaired and restored into a usable state from which normal processing can resume. Even if the more client systems are appended to the system, the stable databases can be guaranteed. And the proposed recovery technique provides the better performance compared with the conventional techniques.

2. System Configurations and Considerations

The switching systems[4] which employ client-server architecture have several considerations in designing the database recovery technique. Figure 1 shows the basic layout of the switching system. Only the server system is physically connected to a disk. All considerable features are originated form the system configurations and the specialties of the switching systems such as the space restriction and the time limitation.

The two basic operations of the DBMS regarding the recovery are the loading and the backup. Note that the disk operations of the whole systems are performed by the server processor.

• At the system initialization, the server processor sequentially loads the contents in disk into the main memory for each server and client system. On the system failure also, the whole system loading is performed.
• While the system running, the logs for the updated database in all systems should be reflected on the disk.

Because of the characteristics of the switching systems, the following considerations should be reflected in designing the recovery technique.

• The volume of the loaded databases should be minimized. In the initialization and the recovery time, most real-time systems limit the total loading time of the whole system to several mi nutes.
• The size and the number of logs should be minimized to reduce the disk I/O requests at run time.
• The allocated volatile memory is restricted to the small size. And the stable memory which simplifies the recovery mechanisms cannot be provided.

3. Real-Time Recovery Technique

The basic paradigm of the proposed recovery technique is that the recovery technique of index structures[5](we abbreviate index structures to indices) should be designed differently from that of the data part(In this paper, data part means the whole database except indices) to tackle the constraints of the systems. To implement the technique, a specific Real-Time Recovery Manager(RTRM) is suggested in each system. It constructs all the indices at the initial loading and the recovery of the system. It is also responsible for recovering the indices while transaction processing.

The RTRM consists of two controllers(Figure 2), Queue Controller(QC) and Index Controller(IC). The QC controls a circular queue, named Rollback Queue(RQ), in the critical section for the exclusive accesses. The RQ maintains the information of the indices to be recovered. The IC constructs all the indices just after the initial loading and it recovers the corrupted indices(in the RQ) by reconstructing them during transaction processing.

The technique maintains the large amount of indices on the main memory only. Only the data part is loaded form the disk. The index-construction is performed by the RTRM, particularly the IC, using the loaded data part. The loading on a client system and the constructing indices on the other client systems are processed in parallel. Due to the parallelism, DBMS contributes to obey the time constraint for the loading and the recovery.

While transaction processing, The RTRM performs the rollback for indices. The time and the memory consuming logging for indices is obviated. When a transaction aborts(Figure 3), the DBMS performs the rollback of the data part using undo logs(arrow(1)). Then the DBMS sends the information of the indices to be recovered to the RTRM(arrow(2)). The RTRM performs rollback of the indices(arrow(3)). The DBMS does not perform the rollback for the indices by itself. Thus, the proposed technique enhances the transaction throughput.

In the switching systems with the client-server architecture, it is advisable for every client processor to share the disk I/O overhead on the server processor.

4. Conclusions

The recovery technique which reinstates the databases is an important issue in real-time DBMS, because the system crash erases the whole databases in main memory. The proposed recovery technique ensures the reliable databases and relieves the systems from the recovery problems. The advantages of the recovery method are as follows:

• Hiring the RTRM in each system lightens the overhead on the server processor significantly and offers an effective recovery for databases.
• Adopting the parallel recovery mechanism contributes to obey the time limitations for the system recovery.
• Reducing the processing time for transactions by obviating the logs for indices makes the more transactions meet their deadlines in real-time systems.

References

1 M. H. Eich, "A classification and comparison of main memory database recovery techniques," In Proc. Intl. Conf. on Data Engineering, pp. 332-339, 1987.

 

2 J. Gray, "The recovery manager of system R," ACM Computing Surveys, vol.13, no.2, June 1981.

 

3 H. V. Jagadish, "Recovering from main-memory lapses," In Proc. 19th Int. Conf. on Very Large Data Bases, pp. 391-404, 1993.

 

4 Y. B. Kim, "An architecture of scalable atm switching system and its call processing capacity estimation", ETRI Journal, vol. 18, no. 3, pp. 107-124, Oct. 1996.

 

5 T. J. Lehman, "A study of index structures for main memory database management systems," In Proc. 12th Int. Conf. on Very Large Data Bases, pp. 294-303, 1986.