SSP

Description

The Sustained System Performance (SSP) an aggregate, un-weighted measure of computational capability relevant to achievable scientific work.  It is used as a way of measuring and reporting 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, # of nodes in the system.  The processing rates, generally expressed in units of GFlops per seecond, are calculated using a floating-point operation count predetermined on a reference system by NERSC.  Thus, the only input required for calculating the SSP consists of the times to execute the benchmark programs. 

The SSP can be summed over a given time period to provide the aggregate computational capability of a system during that entire time.  This can be helpful when system performance changes due to component upgrades or due to system size and it becomes possible to evaluate systems delivered in phases or those in which in-place upgrades are done.  Such a calculation is obtained by summing over all time periods, the product of the SSP times the time period for which that SSP was valid and the result is expressed in units of TF- (or PF-) years.

A sample SSP calculation for a hypothetical workload on a hypothetical system appears below and a spreadsheet carrying out this calculation is available for download.  In this spreadsheet, items in green are supplied by NERSC and are not to be changed; items in pink are to be supplied by the entity running the codes; and items in blue are the calculated values.

More detailed information on the SSP and its use in workload characterization related to previous NERSC acquisitions, is also available. 

Sample SSP Computation

Reference Timings and Counts

The following values are from the NERSC Hopper (Cray XE6) system.  The table will be updated as more data become available.