For its analysis, ButlerBloor Ltd took 11 parallel databases and 14 hardware offerings, and examined each against eight key criteria: market presence and continuity; complex query; simple query; update; hybrid workload; performance accessibility; and portability. In symmetric multiprocessing implementations, Informix and Oracle both figured strongly in the performance categories, with Informix having the edge in the area of complex query, because of what the report considers the careful design of features such as pipeline parallelism. Oracle7 is let down by its limited pipeline parallelism, with only two layers able to be active at any one time. This means that with complex queries, there can be a requirement for large amounts of temporary storage, leading to an increase in disk and a slow down in the pipeline. Version 7.1 stores to disk via the database cache, Version 7.2 uses direct access to disk, significantly improving performance, the report concludes. Another problem is that until Version 7.3, Oracle does not use hybrid hash joins, which when joining tables increases speed. The third weakness is in the slight limitation that Oracle has in the ability to parallelise all queries. While the authors admit that the ability to parallelise small queries is one of the major deficiencies of shared-nothing databases, it is not something from which Oracle suffers. However, Oracle will parallelise a query only if there is at least one full table scan in it. This is required to get a wave of data going up the query pipeline.
Came bottom
Once this has occurred, other components can occur in parallel. Adabas was also highly rated for complex queries, as was Red Brick Warehouse, where it provided impressive indexing capability although its ability to update – a by-product of its target market – is questioned. Oracle Rdb and CA-OpenIngres came bottom in the report’s complex query rating scheme. Neither can parallelise joins, the report concludes. Resource allocation and query optimisation are the key features here. Several products did not have query optimisers and used serial optimisers that know nothing about parallelism, to find a serial execution plan. This means the user may have one processor or 1,000 processors, and the optimiser would find the same query execution plan. Indeed, CA-OpenIngres V1.1 is generally considered poor, lacking effective massively parallel processing coherency control and no-row level locking. In referring to Ingres, the report describes the lead Ingres once had in the relational world, the subsequent cash flow crisis, and ASK Group buy-up as truly lamentable. Computer Associates International Inc said it has already identified [the] weaknesses and will address them in future updates. The company agrees with Sybase about the subjective nature of the comparison (CI No 2,789). The report categorised applications as retrieval-dominated – data warehouse, and update dominated – transaction processing – without touching on such issues as activity location, it points out. Important questions over fat-client, thin server, thin-client and fat server were ignored.
Ray Hegarty concludes his in-depth look at ButlerBloor Ltd’s in-depth 535-page report Parallel Database Technology: An Evaluation And Comparison Of Scalable Systems; part one was in CI No 2,789
Theory is one thing, real-world complex data is quite another, it insists. Sybase MPP does well in pre-planned queries where it has available different partitioning strategies. Red Brick Warehouse is given the thumbs up for ad-hoc symmetric multiprocessing indexing facilities. With ‘simple queries’ – a query with few joins and high specificity – Oracle Rdb came out tops in both the symmetric multiprocessing and massively parallel processing category, largely because it is tightly coupled with operating system mechanisms rather than relying on threads to be fired up. In the symmetric multiprocessing category a whole bunch of products were hard to order including Adabas, Informix, CA-Open Ingres and Oracle7. In the h
ardware arena the report evaluated scalability of performance, hardware configurability, availability, software configurability and management and support. In general, hardware was considered sound, with the majority of the systems able to do the job. Although symmetric multiprocessing machines cannot scale as high as the massively parallel processing architectures, other comparison criteria, such as manageability and hardware and software configurability, contributed to the symmetric multiprocessing boxes scoring higher marks than their massively parallel cousins. The report notes that symmetric multiprocessing and massively parallel systems are undergoing something of an evolution at present, and as parallel hardware technology continues to mature, it expects both hardware and software to converge into a more common architecture. The scalability category had Pyramid Technology Corp’s RM1000 and the Convex Computer Corp Exemplar scoring highest, with IBM Corp’s RS/6000 SP2 and Meiko Scientific Ltd’s CS2 also highly rated. Hardware configurability found Digital Equipment Corp as leader, with its AlphaServer 8400 system, followed by AT&T Corp, Convex, Cray Research Inc, Silicon Graphics Inc and Sequent Computer Systems Inc, all of which were also rated highly.
Relative maturity
Content from our partners
High-availability was dominated by Tandem and DEC with AT&T, Cray, IBM, Meiko and Pyramid offering strong capability. Software configurability was led by the SP2, followed by Cray, DEC, ICL Plc and Sequent. The management of the machine was best dealt with by Tandem and White Cross Systems Ltd, followed by Sequent but in general the report concludes that IBM’s SP2 is the system to beat in all areas of comparison, but without ever quite managing to shine brightest in any one. Symmetric multiprocessing machines performed well, reflecting the relative maturity of the technology. DEC’s 8400 clusters also got the thumbs up, with AT&T’s 5100 considered a reliable system, although rated less scalable than might be expected. ICL’s Goldrush and Tandem do the job, although the report advises against investing in these architectures until new generation products arrive. Pyramid’s RM1000 is also rated highly overall, although the report again recommends ensuring the machine can be configured appropriately. Symmetric multiprocessors that still managed to squeeze into best overall rank included the Cray CS6400 and Silicon Graphics’s Challenge, with Sequent considered a reasonable engine, despite its lower performance. CS6400 highlights include the system bus, input-output, fault-tolerance, large disk capacity and its Sparc-Solaris compatibility, while Challenger offered a fast bus, a powerful R8000 processor and easy configuration.
