The Sustained System Performance (SSP) is an aggregate, un-weighted measure of computational capability relevant to achievable scientific work. It is used as a way of measuring, reporting, and projecting the performance of a given system using a set of benchmark programs that represent a workload. Mathematically, it is expressed as the geometric mean of the processing rates of the benchmark programs multiplied by N, the number of nodes in the system. The processing rates, expressed here in units of TFlops per seecond, are calculated using a floating-point operation count predetermined on a reference system by NERSC. Thus, the inputs required for calculating the SSP consist of the times to execute the benchmark programs, the number of nodes on which they were executed, and the total size of the system for which the SSP is to be computed.
More detailed information on the SSP and its use in workload characterization related to previous NERSC acquisitions is also available.
The reference SSP calculation for the Trinity / NERSC-8 benchmark suite on the NERSC CRAY XE6 "Hopper" system appears below. In the table below, items in green are supplied by ACES / NERSC and would not be changed in a vendor-supplied SSP calculation. Items in pink are the data that would be supplied by a vendor and items in blue are the calculated values. The problem sizes are the "large" cases and the run conditions used here are MPI only. The values shown below are approximate; see the spreadsheet for more precise data. Make sure you read the Run Rules document to understand how these values are calculated.
A worksheet carrying out this calculation will be included in the Trinity / NERSC-8 Benchmark Results Spreadsheet. The spreadsheet will also include a sample SSP calculation for a hypothetical heterogeneous system containing two types of nodes. The Trinity / NERSC-8 Benchmark Results Spreadsheet is now available (Excel97-2004 workbook, xls).
Note well: The SSP is calculated using these reference TFLOP counts along with a time measured on some other system. The objective is NOT to measure the FLOP count on the other system, just the time.
|Total Nodes in System||6384||Results|
|Reference TFLOP Count
(Measured by NERSC; do not change)
|Reference # of Iterations||# of Nodes Used in Running Benchmark||
Measured Benchmark Wall-clock Time
|# of Iterations||Rate (TF/s per Node)|
|SSP (in TFLOPS/s):||52.1|
As noted above, the SSP is calculated using the "large" test problems provided with the codes. For reference purposes only, the floating-point operation counts for the "small" test problem sizes are also available here.