In the week that saw the launch of DEC’s new DECtp transaction processing environment, Belmont-based Oracle Corp decided to add further confusion to the current benchmark brouhaha by releasing the record busting Version 6.0 of its relational database. Clearly captivated by claims of being in the fastest growing segment in the world’s fastest growing market, […]
In the week that saw the launch of DEC’s new DECtp transaction processing environment, Belmont-based Oracle Corp decided to add further confusion to the current benchmark brouhaha by releasing the record busting Version 6.0 of its relational database. Clearly captivated by claims of being in the fastest growing segment in the world’s fastest growing market, the company chose to describe the latest release as the world’s fastest relational database – and proceeded to expend considerable amounts of time and energy justifying the tag. Describing the product as the solution to the great database dichotomy – the good performance low productivity of the heaving, wheezy old database versus the greater productivity-lower performance of the existing relational database products – Oracle argued that Version 6 had set a new series of world record performance levels in all the major transaction processing environments.
Cost-performance On an IBM 3090 mainframe running under MVS, for example, the company claimed processing speeds of 265 tps, clearly knock-ing the Tandem NonStop SQL record of 241 tps from the top spot. In a DEC environment, Oracle claims Version 6 clocked up 49 tps on a VAX 6240, beating the Sybase 29 tps track record, achieved on the more powerful 8800/8700 VAXcluster. Oracle also claims the all time Unix record, registering speeds of 110 tps on a large Sequent supermicro, eclipsing even the 104 tps of Sequent’s experimental Silver Bullet prototype. Equally challenging are the company’s cost-performance claims. With dollars per transaction figures for Version 6.0 ranging from $7,000 for the Sequent box, to $47,000 on the Amdahl 5890-600E, Oracle’s product certainly appears a cheaper bet than rival Tandem’s NonStop $48,000 on the 1601 – or IBM’s quoted $57,000 for DB2 on a 3090 600E. Oracle appeared at pains to reinforce the validity of its performance achievements by offering an in-depth explanation of the TP1 benchmark. The TP1 test is modelled on what happens when one customer deposits money in an account – essentially three updates and one insert – and lays down a number of testing specifications. According to Oracle, however, all kinds of tricks can be employed to make performance results appear more impressive than they really are, with size of database, choice of languages, the use – or not – of multi-processors and the way in which the percentage of transactions meeting the sub one second response time target is measured being the most obvious criteria open to manipulation.
Oracle’s clear conscience is based upon its use – as stipulated of a fully scaled, 5m row database, SQL DBMS and C non proprietary application language, a four-way multi-processor, and the auditing of the 99.9% sub second response rate by the reputable Codd and Date. The company explained the estimated five to ten times performance improvement over Version 5.0 by a series of technical feats, which also enhance the product’s real world application credibility. The aim of the three year Version 6 development programme, explained Oracle’s David Martin, was to provide simultaneous support for both transaction processing and ad hoc managerial type enquiries. Key and unique feature to this end is the development of row-level locking – a sophisticated mechanism that removes the most common of database bottlenecks by ensuring that when a data row is being changed, or reserved for subsequent modification, only that particular row is actually locked. Competing products currently use page level locking, blocking other users from access to a whole page of data at a time. Similarly, the introduction of a multi-version read consistency control technique ensures that, by automatically reading a snapshot of old data, Version 6.0 users are always able to query or report without delay. The elimination of the last real world bottleneck comes in the shape of the no-wait sequence number generator, which presents every user with a new order number as and when they access the database. Normally, order numbers are redistributed from user to user, wi
th number release dependent upon the completion of the first user’s operation. Oracle admits one disadvantage – namely that when users change their minds, gaps are created – but argues that this is outweighed by the obvious advantages. The product’s high processing output – and portability – is also ensured via the its multi-server architecture. Essentially, the fifteen database servers within each central processing unit can be spread between users, allowing performance levels to be maintained as loads change.
Fast Commit Tests run on the DEC VAX 6240 show near linear scalability with the introduction of additional processors; the 11 tps recorded on a single processor increases from 21.7 tps on two processors, to reach 32.0 tps on three, and 42.0 tps with the maximum four. Finally, Oracle claims that it has substantially reduced input output contention – the downfall of Version 5.0 – with the introduction of fast commit – one disk write to perform one or more transactions – and piggy backing, allowing data to assemble and stay in memory for as long as possible before being moved as a whole page. As far as fault-tolerance is concerned, the product now offers on-line database administration facilities, comprising on-line back-up and recovery, and provision to perform functions without shutting the system down and interrupting users. Operations continue on other disks if one disk fails, and a failed disk can be brought back into service by the administrator if it isn’t too far gone; mirror disks are also supported by Oracle if the operating system supports them – VAX/VMS does. If a processor fails in a loosely coupled multiprocessor, Oracle supports transfer of the workload over to another CPU in the cluster running Oracle. And if a network or network node fails, operations automatically continue to function on a distributed system. Version 6 is in limited distribution release, with full production anticipated in three months time. OEM deals have already been signed with Sequent, Prime Computer, Harris Corp, Arix Corp and Control Data. Oracle is now concentrating on developing releases for VM/CMS, Sun Microsystems, Pyramid Technology, Hewlett-Packard MPE, Data General AOS/VS, Apple Macintosh, and for MS-DOS and OS/2.