OTHER USER OPTIONS

The user has other options with FOCUS that help operations. A user can sign on to a user ter-minal and access FOCUS. The user can view the job queues and get a good idea as to the status of his jobs. Another feature is the touch tone tele-phone call. The user calls a certain teletele-phone number and a digitized voice will answer giving the user instructions to touch tone in his user number. FOCUS will then scan the queues for the user's number and respond with the status of his jobs.

CONCLUSION

FOCUS made Computer Operations more efficient. We are now able to operate the facility with fewer people, yet our throughput continues to increase. Before FOCUS, we were using five operators on each shift for a total of fifteen opera-tors to operate five CRAYs. At present, we are using two operators per shift on FOCUS. Even if we add more worker computers to the ICN, the FOCUS staff will not increase.

The user organizations are responsible for the allocations and the scheduling is accomplished by using a centralized control machine. The primary advantage of centralized scheduling and control is that several worker computers can be scheduled, allocated, and viewed as a single production resource.

MULTITASKING PERFORMANCE WORKSHOP SUMMARY

Ann Cowley

National Center for Atmospheric ~esearch

Boulder, CO

Three papers were presented in the workshop. The abstracts are included here, and the papers by David and Dent are included in their entirety.

Koskela's paper was not submitted for publication here.

MULTITASKING THE WEATHER David Dent - ECMWF

The ECMWF Model uses both cpus of a CRAY X-MP/22.

Performance figures will be presented together with measurements of overheads and inefficien-cies. The repercussions of moving to a CRAY X-MP/48 will also be discussed.

VECTOR USE AND CONTENTION MEASUREMENTS Rebecca Koskela - LANL

Performance measurements for parallel and vector processing are reported for the CRAY X-MP super-computers at Los Alamos National Laboratory. The measurements are made with the CRAY hardware per-formance monitor. Three kinds of measurements are made: (1) we measure the percentage of vec-tor instructions executed system-wide, (2) for parallel processing, we measure the amount of memory contention in the CRAY X-MP shared memory architecture for 2, 3, and 4 processors, (3) we also measure the percentage of time a processor is blocked waiting to execute in the shared operating system because another processor is executir.g in it.

CMTS - A CRAY MULTITASKING SIMULATOR Jacques David - CEA-Limeil

CMTS is a CRAY Multit~sking Simulator which can run on CRAY-l or X with CFT 1.10/COS 1.11 and later releases (ALLOC=STATIC), or on CYBER

(NOS/BE - NOS-SCOPE) systems. It can be used for testing and debugging multltasked applications and gpt hering various statistics

(Locks/Events/Speed-up ••. ).

MULTITASKING THE HEATHER

David Dent

European Centre for Medium-Range Weather Forecasts London, England tim-ings which indicate where inefficiencies exist.

HISTORY

The present production model has been developed over a number of years and is used both for research and operational forecasting. The model uses spectral techniques and covers the complete globe. It consists of about 100,000 lines of

ECMWF CRAY-XMP CONFIGURATION

From the point of view of the spectral model, the principal characteristics of the CRAY-X2200 installed at ECMWF are:

80 Mwords/sec memory to SSD transfer rate COMPUTER RESOURCES USED BY THE SPECTRAL MODEL At resolution T106, the single-tasking model requires: pro-tected against simultaneous execution.

GENERAL STRUCTURE

The model is organized into 2 scans over the data, as shown in Figure 1. Within each scan, there is a loop over all latitude rows (160 for the T106 resolution). Between scans is a smaller area of computation associated with diffusion and semi-implicit calculations. The loop over time steps is repeated 960 times for a 10-day fore-cast. However, every 12 steps, significant

addi-A multitasking version of an application requJres while the equivalent antisymmetric part is gen-erated by the appropriate southern row. Both components are combined in different ways to pro-vide contributions to the legendre transform. By computing one northern row and one southern row simultaneously, not only is the memory require-ment minimized, but also the legendre computation is performed efficiently.

Part of the diffusion calculation is also multi-tasked and Scan 2 can be computed 2 rows at a time (see Figure 2).

There remain some relatively small parts of the code which are computed in singletasking mode.

The memory requirements for this multi-tasking strategy are 1.8 Mwords. Note that alternative strategies are, of course, possible. However, subtask structures which may be preferred for optimizing reasons require either more central memory or additional SSD. creation and post-processing of history data.

Since the above timings are very simple and made at the very highest level, they tell nothing about the behavior of individual tasks within the model. Currently, there is no support within the Cray multi-tasking library for obtaining detailed timings. Consequently, all the following timings were obtained by inserting C0de into the model at

For TSKSTART, three distinctly different times are observed as follows:

40 milliseconds for one case only 0.4 milliseconds for 96% of all TSKSTARTs 0.04 milliseconds for 4% of all TSKSTARTs The expensive start corresponds to the very first TSKSTART in the complete application, when addi-tional memory has to be requested from the

and was therefore not held up in the the physical parameterization. Convection and condensation calculations are affected by synop-tic conditions and will therefore vary in space efficiency reasons their effect must be minim-ized. Currently this is achieved by splitting

elapsed seconds/time step speedup

processor execution is due to insufficient memory banks on the X22, where the CP speed is retarded