Pyramid Technology Corp vice-president of marketing Boyd Pearce claims that the Reliant RM1000 parallel system – due to take its bow at CeBit this month (CI No 2,611), will easily surpass the power and scope of even IBM Corp’s most powerful Sysplex mainframe clusters. The Reliant RM1000, he says, will scale up to hundreds of processing nodes and offer storage capacity of greater than 10Tb of on-line information. It will also enable the combination of symmetric multiprocessing and massively parallel processing technology on the same machine, and enable close integration with current Pyramid Nile systems. The RM1000 uses a modular approach, with each cabinet including up to six MIPS R4400 processors and up to 24 disk drives, with associated power s upplies and fans. Individual modules are slotted in via an internal backplane that provides a 50M-bytes per second bi-directional connection between processors (four-way vertical and horizontal per processor, totalling 200M-bytes per second) and Fast & Wide 20M-bytes per second SCSI backplane for disks.

Co-ordinate mesh

The same infrastructure has been used in Nile machines for the last 15 months, which is why integration and upgrades from the Nile, direct to the backplane, are possible. Processors each have their own local memory and communicate through messaging. They are connected in a self-routing X-Y co-ordinate mesh with each processor connected to both the storage and interprocess communications channels. Additional cabinets can be stacked on top or alongside each other. In theory, up to 4,000 processors could be connected together, though the resulting system would be too expensive for anyone to afford. But, says Pearce, there has been interest in systems with hundreds of processors, up to around 600 is practical, and benchmarks of machines with ‘dozens’ of processors have already been carried out, according to Pyramid. Software is crucial to massively parallel systems, and Pyramid has a new version of its Unix operating system, and the results of a year-long collaboration with Oracle Corp on Oracle Parallel Server up and running on the systems. The operating system retains binary compatibility with Pyramid’s existing Unix implementation, although optimisation work is needed on applications to take full advantage. Pyramid says some 20 software partners have already tested their applications, and it should have 20 more ready by the launch. It hopes that one result of its acquisition by Siemens Nixdorf Informationssysteme AG will be a closer relationship with the German software house SAP AG, a relationship that Pyramid has previously found difficult to manage. It now hopes to extend the reach of SAP’s integrated financial, management and manufacturing software beyond mainland Europe to the UK and US, where Pyramid’s sales force is at its strongest.

By John Abbott

It says it already has SAP’s R/3 running on the new system as an Oracle parallel cluster, and that it can partition off the application processing from the database processing onto different configurations of processors. Unlike IBM Corp’s rival SP/2, which Pearce dismisses as a LAN in a can, the Reliant RM1000, Pyramid insists, is not to be viewed as just a bunch of workstations hooked up to the network. Built into the operating system is a virtual storage subsystem, through which all processors have access to and can see all of the storage at any time, giving a single point of control. Also included is high-level partitioning of sets of processors for applications and user communications. This, and the fact that different parts of the machine can be divided into uniprocessor, symmetrical multiprocessing or massively parallel processing partitions, makes the machine suitable for mixed workloads: decision support, data warehousing and business transaction processing can all be allocated to the architecture most suitable. From Pyramid’s existing Reliant operating system offering comes advanced clustering and fault-tolerant features such as configurable RAID, mirroring, st

riping and redundancy, and the Veritas logical volume manager and file system manager. Pyramid views ICL Plc’s Goldrush as the closest Unix-based competitor to the RM1000, ironic given ICL’s partnership with Pyramid on Nile systems. But Pearce claims Goldrush is focused on data warehousing, and is not suitable for the mixed workloads that can run on the Reliant. ICL says that it has focused on multiple database support, Oracle, Ingres, Informix, Adabas, and connection to a variety of front-end systems, offering Fibre Distributed Data Interface rather than Ethernet for fast communications. The RM1000 offers only Oracle Parallel Server from day one, but Pearce says Pyramid is already working on an Informix version 8 implementation for later this year, and is also talking to Sybase Inc, Software AG and Computer Associates International Inc. Some observers claim that while the RM1000 and Nile will be a strong combination sale, RM1000 alone looks somewhat weaker, without full scale systems management of its own. That could make it an attractive to existing Oracle on Pyramid users, but harder to sell beyond that base. The system most comparable to the RM1000 is the currently mythical Teradata-derived AT&T Global Information Solutions 3700 which promised, but did not deliver, a symmetric multiprocessing front-end system tightly coupled to a back-end massively parallel processor. The maximum number of processors is tied to the amount of disk storage available. Today’s largest RM1000 system has 1Tb storage, with 2Tb possible shortly.

Keep pace

While databases in excess of 10Tb are theoretically practical on the RM1000, we are unlikely to see them any time soon. As Peter Slavid of ICL asks Exactly who has a 10Tb Oracle database? Slavid says that a 64 element Goldrush (each element having two processors) can comfortably accommodate 1.5Tb of storage at the moment. By April, when 4Gb disks become available, that will rise to 2.5Tb, and by the end of the year 8Tb systems will be possible. According to Slavid, this rate of change is just about as fast as Oracle can keep pace with. There are now about 10 Goldrush sites in the UK, and a few more in mainland Europe and the US. They are running either Oracle or Ingres, although the first Informix and Adabas systems are due to go live soon. Sof tware AG’s Adabas will be an interesting case, says Slavid, because they actually have databases in excess of 10Tb out in the field – but not under Unix. Out in limited quantities by the end of March, the Reliant RM 1000 goes into full production in the second quarter. There are currently six around the world in beta test. Pyramid says it hopes to sell between $30 to $50m worth of systems this year, for 3.5% of the $1,000m market. Prices start at $200,000 for four CPUs, 16Gb disk.