Relational DBs rev up for high-end TP - transaction processing in client/server environments tests the limits of relational technology - includes a related article on whether the two-tier client/server architecture is sufficient for large-scale online transaction processing

Software Magazine, Oct, 1995 by Barbara Francett

Transaction processing in client/server environments has tested the limits of relational technology. TP monitors help, but the Unix players add features to go it alone.

High-end transaction processing isn't just for mainframes anymore. Increasingly, organizations are migrating legacy systems to enterprise client/server environments to gain a degree of flexibility and accessibility their old systems could supply.

As they do, they must grapple with the need for new levels of performance, in addition to scalability, standards and management issues. The migration from mainframes to open systems for high-end online transaction processing (OLTP) applications requires scalability, to support up to thousands of users; performance terms of both response time and throughput; availability, or fault tolerance; and systems management. Mainframe-based OLTP has exhibited such characteristics for years; now, these capabilities must be available in client/server environments for large-scale OLTP applications to run successfully.

Do Unix-based relational DBMSs have the right stuff? Several relational DBMS vendors are revamping their products to convince users that they do.

Informix Software Inc., Menlo Park, Calif., for example, addresses these issues with the Dynamic Scalable Architecture (DSA), its rewrite of the Informix-OnLine

DBMS to support large numbers of users and high performance with workload partitioning and load balancing. Fault-tolerant features include cluster failover, which allows transactions to shift to another node without shutting the system down. For systems management, Informix will provide tools from Tivoli Systems Inc., Austin, Texas, with the forthcoming Version 8.

Informix V.8 has started shipping to controlled sites and will be generally available early next year, according to Dave Watson, marketing manager for database servers and conectivity. V.8 will extend beyond support for symmetric multiprocessing systems (SMP) to support for clusters and massively parallel systems from IBM, Hewlett-Packard and Pyramid.

Such high-powered platforms will further encourage users to move their high-end OLTP applications down to client/server systems, Watson said. His prediction: "In five years, a whole new generation of systems running on proprietary mainframes will go onto open systems."

In July, IBM announced Version 2 of its DB2 Common Server, a portable implementation of DB2 for AIX that runs on AIX, HP-UX, OS/2, Solaris and, soon, Windows NT. Clients supported include OS/2, Windows, DOS, HP-UX and Solaris. Using the DRDA (Distributed Relational Database Architecture) protocol for communications, DB2 Common Server lets any supported client machine hook to any supported server.

"Now, transaction processing users have access to all the resources in the company," said James Hamilton, lead architect for DB2 Common Server at IBM's Software Solutions Division, Somers, N.Y. Version 2 represents IBM's biggest database release to date in terms of code and functionality delivered, said Hamilton.

Version 2 replaces its predecessor's cost-based optimizer with the two-phase Starburst optimizer, which includes a query rewrite component as well as the standard cost-based capability. "We don't want the user to write SQL and worry about how it is written. The two-phase optimizer takes bad SQL and produces good normalized SQL," Hamilton said.

With Version 2, IBM has completely overhauled the physical storage model and improved such utilities as backup/restore. The company has also added a fast load utility -- capable of handling 2Gb of data per hour -- that directly formats pages and routes them to disk. "The focus is on scalability, in terms of both data size and SMP exploitation," Hamilton said.

The result of these improvements is improved performance, Hamilton said. TPC-C benchmarks run on the single-processor RS/6000 Model R24 resulted in 1,470 transactions per minute (tpm). On the eight-processor SMP RS/6000 Model J30 the outcome was 3,119 tpm.

Enter the `Middleman'

The issue of scalability extends to more than the number of computing engines supported, said Dan Lahl, director of server product marketing at Sybase Inc., Emeryville, Calif. Also important are how fast backup and 1/0 can be performed, as well as how many users can be supported, he said.

Sybase SQL Server 11 partitions and optimizes memory for OLTP applications to increase their performance, Lahl said. A logical memory manager lets users carve up memory space to use it more effectively.

Users can also set up memory partitions, putting critical tables inside memory and making them memory-resident. Data partitioning lets users slice up tables into segments for loading and unloading. Production shipments for SQL Server 11 are slated to begin Q4/95.

For many users, the move to large-scale, Unix-based OLTP requires a "middleman" -- a transaction processing monitor -- and a three-tier architecture. In a two-tier client/server architecture, the application runs on a client and the database runs on a server. A three-tier architecture employs a desktop client, one or more application servers, and a database server. Only the GUI component runs on the desktop; the rest -- business logic and database access -- runs on the servers.