The collaborative PowerParallel box…
Buried within IBM announcement of its PowerParallel multi-processing RISC system last month (CI No 2,113) was the news that its parallel processing capabilities will opened up to other vendors’ machines. Processing work will be able to be farmed out to networked Unix-based machines from the likes of Sun Microsystems Inc, Silicon Graphics Inc and Hewlett-Packard Co, even IBM mainframes running AIX/ESA will be able to join the melting-pot. The plan relies on a clever piece of software called LoadLeveler, which was announced alongside the PowerParallel machines. As its name implies, LoadLeveler enables users to establish which processors should work on which job and whether they are working co-operatively, or concentrating on their own discrete tasks. Though the software was announced with the PowerParallel machine, it already has the ability to run processes on remote IBM AIX workstations. A quick look at the architecture of the first PowerParallel machine (called SP1 by IBM) show why this is relatively easy. The basic building block is a single frame containing eight or 16 RISC processors and up to four frames can be joined together. Surprisingly though, in the normal configuration all of these parallel processors are linked by a standard Ethernet local area network. There is little logical difference, then between 16 workstations on a local network and a chassis with 16 processors. Under a heavy parallel processing load, the whole thing will tend to slow down as the 10Mbps links clog, so at the SP1’s launch, the company concentrated on its new, high speed packet switch which can replace the Ethernet. Using this – it is optional – the machine’s processing capacity grows linearly as new processors are added, from a claimed one 1 GFLOPS to 8 GLOPS with the maximum of 64 processors on board. The company says that it intends to support the FDDI, Fibre Channel Standard and Parallel Channel Attachment standards but this will be for connecting the SP1 to file servers and hosts rather than speeding up the local network within the chassis.
…and LoadLeveler to pull it together
LoadLeveler, then, can work in a number of ways: it can schedule batch jobs on groups of standard RS/6000 workstations, while on the SP1 it can schedule batch and parallel batch jobs. It can even schedule parallel jobs across groups of workstations although the company is not hyping this facility, presumably because of the communications bottleneck caused by Ethernet. Each workstation on the network can be set up to decide whether it will accept work being thrown at it, and tells the central manager what resources (memory size, architecture and so forth) that it has available for running jobs. Users with a horror of suddenly finding their workstations grinding to a snail’s pace need not worry since they can stop LoadLeveler from using their machines when they are logged on. Similarly, dedicated compute servers can be configured so that they will run only certain types of jobs, or have a preference for certain types. In order to make best use of the available resources, LoadLeveler needs to be given information about the code that it is trying to run. By way of pay-off, each of the workstations need to set aside 10Mb to 15Mb of disk for log and work files, while the master machine needs to reserve 200Mb. How much memory it takes up depends on the applications being run, comments IBM.
Parallel Environment: Making it work
The SP1 runs AIX/6000 Version 3.2.3 on top of which sits the Parallel Operating Environment. One big attraction is the claim that it standard AIX applications will generally run with out modification. However programs that try to access facilities not available in the box such as graphics adaptors will need tweaking. The programs run within the Parallel Operating Environment, which retains most of the features of the standard AIX/6000 Motif-based interface. The Parallel Operating Environment is just one piece of the bundle of utilities and development tools that IBM has dubbed the AIX Parallel E
nvironment. This contains a collection of tools for the design and development of parallel applications both in Fortran (a nod towards the scientific community) and C. The environment has to be running on at least one processor while the rest of the processors in the cluster need parallel client software installed. The Environment as a whole embraces parallel application programming interfaces that implement message-passing facilities common in distributed memory machines; the PDX Parallel debugger and a performance visualisation and monitoring tool that enables users actually to see how their application is running, and spot bottlenecks. Applications can be developed on a single machine and then run in parallel later. In addition to its own application programming interfaces, IBM is also supporting the Express, Linda and Parallel Virtual Machine interfaces, which should help users bring over existing parallel programs to the new machine. It may even be possible to convert old non-parallel Fortran programs with the help of the Forge 90 tool from Placerville, California-based Applied Parallel Research Inc, which IBM says it is also supporting.
New PowerStations and PowerServers
There were significant extensions to the workstation range, alongside the parallel machine: right at the bottom is the PowerStation M20 a diskless system with a single 33MHz processor. It includes integrated Ethernet adaptor and a SCSI interface, so using the latter, it is possible to add a disk. Higher up in the range comes the 355 the 365 and the 375. These three differ from their predecessors by incorporating a 32Kb instruction cache, four times bigger than previous models. Other than that, the main differences between the models is processor speed: the 355 is a 41MHz machine, the 365 and 375 run at 50MHz and 62MHz respectively.
High Performance Network Server with a little help from Auspex
The IBM 7051 Power Network Dataserver is IBM’s response to the need for a high speed file server to work with these new machines, and was developed with Auspex Systems Inc. It has taken its PowerServer 340R and has bolted on extra communications and data storage options. IBM says the base PowerServer’s scalable asymmetric multiprocessing system, coupled with a maximum of 144Gb of disk and 384Mb of cache can serve workgroups of over 100 users. Some of the added capabilities include Escon optical channel support, data striping across multiple disks and disk mirroring. The former improves performance by decreasing access time, the latter gives the ability to perform on-line back-ups without bringing the system down. The machine acts as a general Network File System-compatible server and should manage around 2,000 Network File System operations per second, though this is dependent on the nature of the application, IBM notes.
