O. M. Lubeck, et. al., "The performance realities of massively parallel processors: a case study", Supercomputing'92, IEEE Computer Society Press, November 1992, pp. 403-412.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....(cf. Table 5) But the disturbing factor here is that even the vector programmers are extremely unrealistic in their expectations. Recent studies have shown that even well ttmed programs are likely to achieve less than 20 of the aggregate peak performance on today s multiprocessor systems [1,4]. Only 18 of all subjects, however, recognized that they would probably get a factor of 25 or less improvement if their parallel application ran on lee processors. There was little difference in vector and serial responses at this low end, although serial users were significantly more likely to ....

Lubeck, O. M., M. L. Simmons, and H. J. Wasserman, "The Performance Realities of Massively Parallel Processors: A Case Study," Proceedings of Supercomputing '92, Nov 1992, pp. 403-412.

....Barton Miller [13] ffl Distributed Performance Monitoring: Methods, Tools and Applications, R. Hofmann et al. [86] ffl A Visualization System for Parallelizing Programs, C. Dow et al. [87] ffl The Performance Realities of Massively Parallel Processors: A Case Study, O. Lubeck et al. [88]. ffl Performance Visualization for Parallel Programs, E. Lusk [89] ffl The RP3 program Visualization Environment, D. Kimelman [90] ffl Visualization of Program Performance on Concurrent Computers, D. Rover et al. [91] 48 ffl Simulation and Visualization Tools for Link based Parallel ....

Olaf Lubeck, Margaret Simmons, and Harvey Wasserman, "The Performance Realities of Massively Parallel Processors: A Case Study", in Proceedings of Supercomputing '92, Minneapolis, November 1992, pp. 403--411.

....(cf. Table 5) But the disturbing factor here is that even the vector programmers are extremely unrealistic in their expectations. Recent studies have shown that even well tuned programs are likely to achieve less than 20 of the aggregate peak performance on today s multiprocessor systems [1,4]. Only 18 of all subjects, however, recognized that they would probably get a factor of 25 or less improvement if their parallel application ran on 100 processors. There was little difference in vector and serial responses at this low end, although serial users were significantly more likely to ....

Lubeck, O. M., M. L. Simmons, and H. J. Wasserman, "The Performance Realities of Massively Parallel Processors: A Case Study," Proceedings of Supercomputing '92, Nov 1992, pp. 403-412.

....the reader is struck by the fact that both the Cray and the CM 2 can be thought of as SIMD processors. A code that has been implemented efficiently on a CM 2 should be directly translatable to a efficient vectorizable code on the Cray. An excellent demonstration of this principle can be found in [10], where three applications that had been programmed and tuned on the CM 2 were ported back to the Cray Y MP with high performance. Thus the reader of [16] is left to wonder why, if the CM 2 code for this application runs so well, did the author not try to adapt the CM 2 code to the X MP Since one ....

O. M. Lubeck, M. L. Simmons and H. J. Wasserman, "The Performance Realities of Massively Parallel Processors: A Case Study", Proceedings of Supercomputing '92, to appear.

....Achieving a significant fraction of this performance would allow the study of many important physical problems (compressible, turbulent flows [PoP92] meso scale weather prediction [DrJ92] ocean circulation, etc. that would otherwise be intractable. Unfortunately, recent published results [LuS92] have suggested that these performance levels are not yet sustainable on complete applications codes. The goal of the work described here is to achieve very high performance on MPPs while retaining portability and ease of programming for real applications codes. The approach we propose is to ....

.... high performance often advertised for MPPs is achieved primarily for small kernels and benchmark programs [BrH91] Recent performance studies comparing current MPPs to conventional vector supercomputers have shown that MPPs sustain a smaller fraction of their peak performance than vector machines [LuS92]. This seems especially true for 1. Woodward first developed several of these ideas when implementing PPM on a Cray using solid state disks. June 24, MPP 3 Fortran P applications requiring frequent global or irregular communication [BaB92] though this result is unsurprising ....

[Article contains additional citation context not shown here]

O. Lubeck, M. Simmons, and H. Wasserman, "The Performance Realities of Massively Parallel Processors," Supercomputing '92, pp. 403-412, Minneapolis, MN, November 1992.

....[20] 23] Achieving a significant fraction of this performance would allow the study of many important physical problems (compressible, turbulent flows [16] meso scale weather prediction [7] ocean circulation, etc. that would otherwise be intractable. Unfortunately, recent published results [11] have suggested that these performance levels are not yet sustainable on complete applications codes. The goal of the work described here is to achieve very high performance on MPPs while retaining portability and ease of programming for real applications codes. 1 This work was supported in part ....

O. Lubeck, M. Simmons, and H. Wasserman, "The Performance Realities of Massively Parallel Processors," Supercomputing '92, pp. 403--412, Minneapolis, MN, November 1992.

....is that most arrays are accessed with a stride equal to the length of the grid. NEUT: A highly vectorizable Monte Carlo neutron transport code. that runs a k calc computation starting with 32K neutrons. NEUT represents a Fortran77 version of Eldon Linnebur s (LANL) Connection Machine Fortran code [12]. SWEEP: SWEEP3D is a three dimensional solver for the time independent, neutral particle transport equation on an orthogonal mesh [7] The first order form of the transport equation is solved by sweeping through the spatial mesh along discrete directions (ordinates) The solution algorithm in ....

O. M. Lubeck, M. L. Simmons, and H. J. Wasserman, "The Performance Realities of Massively Parallel Processors: A Case Study," Proc. Supercomputing `92, IEEE Computer Society Press, 403-411, 1992.

....A typical problem is run on a 100 X 100 mesh for 100 time steps. An important characteristic of the code is that most arrays are accessed with a stride equal to the length of the grid. The executable size is about 5.8 MBytes. NEUT: A highly vectorizable Monte Carlo neutron transport code [8]. Two problem sizes may be run, one starting with 32K neutrons, the other with 64K neutrons. NEUT represents a Fortran77 version of Eldon Linnebur s (LANL Group X 1) Connection Machine Fortran code. The executable size is approximately 34 MBytes. SWEEP: SWEEP3D is a three dimensional solver for ....

O. M. Lubeck, M. L. Simmons, and H. J. Wasserman, "The Performance Realities of Massively Parallel Processors: A Case Study," Proc. Supercomputing `92, IEEE Computer Society Press, 403-411, 1992.

No context found.

O. M. Lubeck, et. al., "The performance realities of massively parallel processors: a case study", Supercomputing'92, IEEE Computer Society Press, November 1992, pp. 403-412.

No context found.

O. M. Lubeck, et. al., "The performance realities of massively parallel processors: a case study", Supercomputing'92, IEEE Computer Society Press, November 1992, pp. 403-412.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC