Home Go Up

NEC SX-8 Benchmark

Machine Description

Model Number SX-8
Number of Processors 8
Memory Installed 64 Gbytes
Memory Used by FMS 51 Gbytes
Operating System SUPER-UX Release 15.1

Problem Description

Code Used FMS Example 11
Matrix Sparsity Full
Matrix Symmetry Symmetric and Nonsymmetric
Data Type 16 Byte Complex
FMS Parameter IALGOR 0
Number of Vectors 1

Results for Complex Symmetric Matrices

Date
Started		 Number of
Equations		 Storage Times (Day::Hr:Min:Sec) Megaflops Parallel
Speedup
Loc. Gbytes CPU Wall I/O Wait CPU Overall
02/11/06 20,000 Mem. 3.7 11:52 1:55 0:00 14,992 93,006 6.2
02/11/06 20,000 Disk 3.7 11:53 1:56 0:00 14,965 91,920 6.1
02/11/06 40,000 Disk 13.8 1:33:55 13:10 0:00 15,145 107,999 7.1
02/11/06 60,000 Disk 33.7 5:13:07 42:27 0:00 15,330 113,082 7.4
02/11/06 80,000 Disk 75.7 12:09:19 1:36:43 0:40 15,601 117,635 7.5
02/11/06 100,000 Disk 106.0 23:37:45 3:05:08 0:51 15,674 120,035 7.7
02/12/06 150,000 Disk 227.2 3::07:19:52 10:15:28 6:43 15,757 121,861 7.7

Results for Complex Nonsymmetric Matrices

Date
Started		 Number of
Equations		 Storage Times (Day::Hr:Min:Sec) Megaflops Parallel
Speedup
Loc. Gbytes CPU Wall I/O Wait CPU Overall
02/12/06 20,000 Mem. 7.5 22:38 3:23 0:00 15,706 105,288 6.7
02/12/06 20,000 Disk 7.5 22:40 3:25 0:00 15,683 104,105 6.6
02/12/06 40,000 Disk 27.5 3:02:33 24:39 0:01 15,583 115,378 7.4
02/12/06 60,000 Disk 67.3 10:14:40 1:21:29 0:02 15,618 117,823 7.5
02/12/06 80,000 Disk 151.5 23:59:57 3:09:04 2:03 15,803 120,366 7.6
02/12/06 100,000 Disk 212.1 1::22:43:57 6:03:39 3:12 15,851 122,223 7.7

NOTES:

  1. CPU Time is the total amount of User and System time used by all processors.
  2. Wall Time is the elapsed time measured on a dedicated machine. This includes the time spent processing and any time spent waiting for I/O to complete.
  3. I/O Wait Time is the total time spent waiting for I/O to complete that is not overlapped by asynchronous I/O. This is included in Wall Time. This benchmark was performed on a machine prior to installation and high performance disks were unavailable, resulting in the large I/O wait time and lower parallel speedup.
  4. Megaflops per Processor is the theoretical number of floating point operations performed, divided by the CPU time. When reduced operation algorithms are used, this is based on the number of floating point operations that would have been performed by the traditional algorithm.
  5. Overall Megaflops is the theoretical number of floating point operations performed, divided by the Wall Time.
  6. Parallel Speedup is overall increase in performance due to parallel processing.

Home Go Up
Copyright © 2005 Multipath Corporation