For a comparison with other parallel supercomputers, the following is a table of the performance of our parallel treecode running a 10 million particle benchmark. All machines are running the same code, with the exception that the Intel i860 machines and the CM-5 have the inner loop coded in assembly language. The code for Loki is entirely in C, and was compiled with gcc 2.7.2. Message passing was accomplished with our own UDP socket library.

Treecode performance

Site	Machine	Procs	Time	Gflops	Mflops/proc
LANL	TMC CM-5	512	140.7	14.06	27.5
Caltech	Intel Paragon	512	144.4	13.70	26.8
NRL	TMC CM-5E	256	171.0	11.57	45.2
Caltech	Intel Delta	512	199.3	10.02	19.6
NAS	IBM SP-2	128	281.9	9.52	74.4
JPL	Cray T3D	256	338.0	7.94	31.0
LANL	TMC CM-5 no vu	256	754.6	2.62	5.1
SC '96	Loki+Hyglac	32	1218	2.19	68.4

Time is wall clock time in seconds, and includes all message passing and load imbalance overheads.

If you are interested in a further description of the algorithm, please see the papers describing the treecode and our NASA HPCC project page.

Back to Loki Home Page