There was a lively attendance of about thirty-five people for the Atlas Cray Users' Meeting held on 30th April. Those who were expecting news of the new supercomputer procurement were disappointed, since it was not yet possible to report any decisions.
Roger Evans gave a brief review of the service provided over the last few months. Hardware and software reliability had been good but the overall availability figures were adversely affected by a loss of two days service due to an electricity supply failure on Christmas Day!
Utilisation of the machine remained consistently high, but job turnaround varied from just acceptable to poor when the machine was most heavily loaded. This was felt to be a major concern in planning the loading on a future supercomputer.
There had been little software development over the past five months, due to a heavy involvement in the procurement and benchmarking of the new supercomputer. System development was now concentrated on a reduction of the local modifications to UNICOS source code needed to provide the current level of service. This should help system reliability in the future, as well as reducing the effort needed to maintain and install future UNICOS releases.
The major development in removing local software from UNICOS would be in the introduction of new accounting software developed by Mike Froggatt (RAL) and Andy Bailey (Cray). The current accounting software relies on an authorization request to the IBM front-end every time an interactive session or batch job is started. This makes maintenance of the single database easy, but means that new work on the Cray cannot start if the IBM is unavailable.
Under the new scheme, Cray's CSA (Cray System Accounting) software will be used to gather usage statistics every few hours. These will then be sent to the IBM and will update the usage database. Software on the IBM will then inform the Cray of accounts that have been overused and they will be disabled.
The major advantage of using this method is that it is an application code running independently of UNICOS and is much easier to maintain. The minor disadvantage is that an account can overspend in the period between a job starting and the updating of accounting information between the Cray and IBM. Users felt that this would be acceptable if the amount of overspend could be limited to a single job.
The account specifying option # QSUB -A account is not a standard feature of NQS but is supported by local code. Neither UNICOS version 7.0 nor the currently-used level of RQS on the UNIX front-end machine recognise this option. We plan to request all users to replace the # QSUB -A instruction with the UNICOS command newacct, which will set the current account in a batch job.
There is a long-standing need for more flexibility in the order in which jobs are started, to allow different patterns of usage from different users. David Gaunt described his work to provide this on the IQS system. A batch job would be piped through a holding queue and would be sent to an execute queue only when there was an execution slot available. The IQS software would periodically inspect the holding queues and select for execution jobs which matched any criteria specified.
The initial requirement is to start jobs in a 'round robin' of user IDs, so that users who submit several batch jobs at once do not block others from using the large job queues. In future we could give users the opportunity to reorder their own jobs, and also, give higher priority to users with very large grants who currently have difficulty in getting their allocated CPU time due to the overload on the machine.
John Gordon gave an update on the applications software available on the UNICOS service. Recent additions include Phoenics / Photon, Dyna3D and MPGS. Restricted user licences have been obtained for Discover and NCAR graphics, these being cases where a general licence was either too expensive or could not be obtained for contractual reasons.
Software currently available for the UNIX front-end service includes the IP and X.25 communications software, both NQS and RQS to control batch jobs on UNICOS, and X windows, currently at release 3 but due to move to X11R4 shortly.
Users were asked to make requests for specific application software for the RS/6000 front-end; a full-featured text editor and a powerful graphics package, such as PVWave, were requested.
Brian Davies gave a brief update on the supercomputer procurement exercise, apologising that it was not yet possible to say anything about the outcome. The operational requirement and benchmarking exercise had been completed, but decisions would not be made until approved by the ABRC sub-committee, and possibly not until after the SERC Council meeting in June. It could be assumed that a new supercomputer would be available to the research community around October 1992.
Brian Davies and Richard Catlow both thanked all the Atlas Cray Users who had helped to prepare the benchmark material.
David Lary from Cambridge presented a short summary of the work of the EASOE (European Arctic Stratospheric Ozone Experiment) project which had recently had high priority access to the Atlas Cray.
Modelling the ozone chemistry of the upper atmosphere is very complicated. Because much of the chemistry is driven by sunlight, it needs a thorough treatment of radiative transfer in the atmosphere, whereas other reactions, which take place on the surface of cloud droplets, need a model of stratospheric cloud formation. The resultant equations are "stiff" and moreover the chemistry determines the temperature structure in the stratosphere and reacts back on the atmospheric dynamics.
The model used is based on the ECMWF global circulation model run at relatively low resolution. Nevertheless the additional chemistry and feedback processes make the model expensive to run and simulations are typically of only one day or so of real time.
Novel methods have been found to initialise the chemical concentrations for the simulations which dramatically improve the usefulness of the models and there is now a very good understanding of the interaction of the temperature, cloud formation, sunlight and atmospheric dynamics. Ozone depletion measured in the Spring of 1992 is noticeably affected by dust and sulphur aerosols from the Mount Pinatuba explosion last year.
Atmospheric chemistry modelling is clearly very important in the current political climate, but with current computing resources it is restricted to modelling very short periods of time compared with the climate modelling programme.
After lunch, Professor Catlow introduced the open discussion session. A lively topic was the allowable over-allocation of computing resources that should be planned for the new national supercomputer. Currently the allocation of time is limited to 130% of the time actually available. The additional factor is to allow for the fact that inevitably there will be some grants that fail to use their awarded computer time for reasons of manpower or changed scientific needs.
Experience on the Atlas Cray has been that at the 130% level of allocation the batch queues are so long that the perceived performance of the supercomputer in terms of job turnaround is very poor. If national supercomputers are to be used sensibly then the turnaround must be improved. The consensus of the meeting was that a reduced over-allocation to 120% should be suggested to the Supercomputer Management Committee.
The effective use of the new supercomputer was also discussed in terms of the step-function that would be introduced to the available resource and how allocation could be managed to cope with the rapid change. Both supplementation and more rapid usage of existing grants were possibilities, and it was also suggested that SERC find a way to award large amounts of time in the first six months to projects that could make immediate advances in important areas.
Discussion of the fate of the current Cray X-MP at RAL was hindered by lack of information on a new machine. Users expressed the view that if the new machine were not a Cray, there would be a clear need to retain the current Cray service, whereas if the procurement resulted in a new Cray at the Atlas Centre then the X-MP could be removed after a suitable overlap period.
