PVM (Parallel Virtual Machine) is a widely available and free library of subroutines to implement parallel processing on a group of similar or heterogeneous computers. PVM version 2.4 is available on the Y-MP at Atlas and PVM is available for most flavours of workstations running Unix (DEC, SGI, IBM RS6000, SUN and others).
The purpose in producing PVM was for users to be able to distribute single applications over a network of workstations and make use of otherwise idle CPU cycles to tackle problems that were too large for a single workstation to perform in a reasonable time. PVM has Fortran and C calls to implement all the usual primitives associated with message passing methods for distributed memory parallel processing. PVM will run over wide area networks such as JIPS and SuperJIPS (the IP service of SuperJANET).
We see the value of PVM on the Cray as twofold: firstly, if you have developed message passing code for other machines it should be fairly easy to translate to the PVM library and run the application on the Cray. The number of processes in PVM is not limited by the number of Cray CPU s and can be decided by the structure of your application. The overhead of the PVM subroutine calls is somewhat more than implementing Cray's own multi-tasking constructs but if you already have the message passing code then you can try it out on the Cray and decide whether the overheads are acceptable. Secondly, and perhaps more importantly, we see PVM as offering a very straightforward route to distributed processing between local workstations and central supercomputers. It is very straightforward to set up the workstation code to start a process on the Y-MP, send a message containing the data for a particularly intensive portion of the calculation and receive a message with the results from the Cray. Typical applications might be to run a molecular dynamics simulation on the Y-MP, with the molecular coordinates being sent to the workstation to give a near real-time display. This would be vastly more efficient in terms of network bandwidth than sending the graphics image over the network using X-windows.
PVM is one of a number of subroutine libraries for parallel processing and the similarity of function but difference of syntax has led to the attempt to generate a common user interface. One such is the macro language PARMACS and this has been chosen for the distribution of the benchmark codes for the planned acquisition by SERC of a major parallel supercomputer. PARMACS has been implemented on the Y-MP for the purpose of establishing a baseline for the benchmarking effort but is generally available to users. PARMACS has the great advantage to users starting from scratch that it caters for the commonly used domain decomposition structures such as regular grids, and automatically generates the message passing calls.
There are man pages on UNICOS for PVM: look under pvmd for the PVM demon that must be run on each participating machine, and pvm_intro for more details on the syntax of the PVM subroutines.
PVM is generally available from its originators at Oak Ridge National Laboratory. If you want to install it on your own workstation, send e-mail to netlib@ornl.gov with the one line message
send index from pvm3
This will give you instructions to get the whole of the code and documentation.
The Computational Fluid Dynamics Community Club is part of the Engineering Applications IT Support Programme. It was founded in late 1989 and now has a registered membership of 500 academics and industrialists representing more than 65 higher educational institutes. Three of its objectives are to increase the awareness of advanced computing and software engineering techniques in the community, to encourage an exchange of views and information, and to promote the exchange of CFD software and data throughout the community. One of the requirements of this Community Club was easier access to commercial CFD software on supercomputers. To satisfy this need we have negotiated special licences with a number of commercial software vendors to allow academic use of their products on the Cray Y-MP8 at the Atlas Centre.
The packages involved are:
The arrangement for the first four packages is that existing academic holders of workstation licences can use the same software on the Cray without obtaining a separate, usually more expensive, Cray licence. The fifth package, FLOW-3D, is freely available to all academic users. Although there are no software costs involved, an Research Council grant is required to cover the CPU time used on the Cray. A small amount of CPU time (5 Y-MP CPU hours over three months) can be obtained through a Pump Priming award in order to evaluate the software. Pump Priming awards can be obtained quickly, but only by those eligible to apply for Grants (tenured year academics at UK HEIs) or Research Council staff.
Readers are reminded that SERC grants which are only for computer time up to a maximum of 250 Y-MP CPU hours can be obtained through the fast track method which is processed far more quickly than the normal bi-annual grant round.
Before the JIPS (JANET IP Service), mail across JANET was sent using a system known as Grey Book Mail, part of the Coloured Book Software suite. Each site tended to have just a few machines capable of sending mail off-site. Mail destined for other sites on JANET could be sent directly using Grey Book Mail; other mail was passed to one of several national mail gateways such as EARN-Gate or NSFnet-Relay.
With the advent of the JIPS, however, the number of machines able to send mail to other sites has greatly increased. Unfortunately, using the JIPS to send mail involves using SMTP (Simple Mail Transfer Protocol), a protocol not supported by the JNT (Joint Network Team). Although SMTP appears to be relatively easy to use, it does pose a number of problems.
Most commonly, mail messages sent via SMTP have invalid "reply-to" fields so that the recipient cannot easily reply to the sender. This happens when the sending host is in the "ac.uk" domain but is not registered in the NRS (Name Registration Scheme). Many Grey Book mailers are not able to handle such messages while others transfer these messages via NSFnet-Relay, the international SMTP/Grey Book gateway, causing overloading of this valuable resource.
If care is taken, however, SMTP may be used under certain circumstances: for example, when sending mail which would otherwise pass through NSFnet-Relay destined for machines outside the UK. SMTP may also be used between machines within the UK provided that all mail header addresses are also valid Grey Book mail addresses, i.e. that the machines concerned are registered in the NRS. Individuals wishing to use SMTP in this way should first consult their local network providers who may be able to offer an alternative.
Some time ago, a policy for JANET was adopted that mail within the UK should progress towards the ISO standard, adopting X.400 as the desired mail transfer mechanism. As part of this commitment, the JNT funded the development of a site mail gateway or "mail hub", capable of handling X.400, SMTP and Grey Book mail. A mail hub is a single transaction point for all mail arriving at or leaving a site or large department and is designed to facilitate efficient and coordinated mail delivery across networks.
As a centrally provided mail service, PP mail hubs have a number of benefits over existing mail systems, both in terms of functionality and in terms of the levels of support available. PP can handle most current mail protocols including X.400 (84 and 88), Grey Book, SMTP, and POP with further developments being driven by the needs and wishes of the UK academic community.
Mail from workstations and PCs can be routed via a local PP mail hub, minimising the administration of mail on those systems and allowing machine names within a site to be hidden from the outside world. For example "user@machine.site.ac.uk" could be changed to "user@site.ac.uk". Returning mail could then be routed to the user's preferred mailbox on site rather than the machine from which the message was originally sent.
PP can also act as a "message store", providing a single mailbox for each user in which to keep mail. This mailbox would be available to the user from whichever workstation or PC was currently in use and would prevent being tied to any one machine for reading mail. This mechanism also eliminates the need for leaving PCs and workstations continually active awaiting incoming mail. Both Microsoft Windows and X Window based Mail User Agents (MUAs) are available - contact your local network providers for more information.
Generally, although it is possible to send mail directly from workstations and PCs, users are well advised to find an alternative such as routing via a local PP mail hub. This requires less effort to configure and relieves the administrator of the workstation or PC from the responsibility of having to manage the rather tangled web of global e-mail. A large number of UK sites are now adopting PP mail hubs so it is likely that one is available for you now.
For more information on the use of SMTP contact your local network providers. They should also be able to advise you on the availability of a local PP mail hub and suitable mail user agents.
SMTP (Simple Mail Transfer Protocol) is an IP (Internet Protocol) based mail protocol using direct connection between originating and destination hosts. SMTP is very insecure and such mail messages are easily forged.
NRS (Name Registration Scheme) is the scheme of tables used by the X.25 Coloured Book Protocols containing the name and address of all the X.25 connected hosts on JANET.
MUA (Mail User Agent) is the system which allows users to manipulate (read, delete, save, etc.) mail messages at their terminal; programs such as Whitemail, ELM and Z-Mail.
At the end of 1992, the SERC approved the purchase of a new supercomputer to be installed at MCC. It will replace the Amdahl VP1200 which will have been in service for five years.
The new supercomputer continues the programme of investment in national high performance computing facilities, overseen by the Advisory Board for the Research Councils and managed by SERC.
The supercomputer will enable new science to be researched in areas where the problems need extremely high performance from programs or application packages which require either very large memory or have inherently long vectors. These occur in a broad range of science and engineering projects sponsored by the UK's Research Councils. It is also likely to attract new types of project and may enable scientists to tackle in this country some of the large projects which could previously only be done on supercomputers in other countries.
The new supercomputer is a Fujitsu VPX240/10, a configuration that has 1 scalar processor, 1 vector processor with a peak performance of 2.5 GigaFLOPS, 1 Gigabyte of main memory (MSU), 1 Gigabyte of secondary memory (SSU), and 44 Gigabytes of disc. The operating system will be UXP/M, Fujitsu's UNIX, which is an implementation of System V Release 4.
The architecture is similar to that of the VP1200 which it replaces, but has 32 times more memory, and is over 4 times faster. The greater speed is due to having both a larger number of arithmetic pipelines and an increase in the clock speed, which is 6.4 nanoseconds on the scalar processor (d. 14 nanoseconds on the VP1200), and 3.2 nanoseconds on the vector processor (cf. 7 nanoseconds on the VP1200).
Delivery is expected in July 1993, with an aim to complete installation and acceptance in time for a service start at the beginning of September 1993. MCC will be assisting researchers to transfer their work to the new service, and the VP1200 is expected to be removed at the start of December 1993.
