NERSCPowering Scientific Discovery Since 1974



STREAM is a simple, synthetic benchmark designed to measure sustainable memory bandwidth (in MB/s) and a corresponding computation rate for four simple vector kernels.


NERSC8-Stream tar file

How to Build the Code

The version of STREAM provided here is the latest STREAM OpenMP enabled benchmark.

Optionally, for GPU architectures, you may choose to run the Scalable Heterogeneous Computing Benchmark Suite (SHOC) Level 1 Triad benchmark. (see See run rules below.

You need to supply a good timer routine. You should compile this program with the highest optimization possible for your particular platform.

The STREAM source code has a parameter, N, that represents the length of the arrays used in the tests. It is important to set N such that the test touches a large enough portion of the memory. The vectors are of type DOUBLE PRECISION and there are three vectors A,B,C of this length. So for 8 byte doubles 3*8*N bytes are required. 

N is hardcoded as a PARAMETER, located
around line 98 of the FORTRAN stream.f source code.
around line 58 of the C stream.c source code 

The STREAM source code allows the use of an offset parameter to separate the three arrays used and you may use any value for this offset although you should state the reason for the specific offset chosen, either as a comment in the source code (which should be returned to NERSC) or elsewhere in your materials.

To Build

You must also return to NERSC either a makefile or script or something similar that shows all the compiler/linker options used for compilations of STREAM. An example makefile and runscript are provided.

Edit the Makefile to use the appropriate compilers and compiler flags.
Type 'make'
This will build both the C and Fortran executables.

Required Runs for Trinity/NERSC-8 

On a multi-core node with "n" hardware cores and M bytes of main memory STREAM must
be executed as follows:

  1. The parameter N adjusted such that the used memory is 0.6*M
  2. Two results must be reported:

a) maximum bandwidth with enough OpenMP threads to utilize all hardware cores and hardware threads (fully packed)
b) bandwidth for a vendor determined minimum number of cores and threads that achieves maximum bandwidth

If results a) and b) are the same, only report a single result. In both cases, report the number of cores and threads used as well as their placement and affinity, if applicable.

Please report memory bandwidth results in the provided spreadsheet.

Verification of Results

The code is self-checking.

The STREAM benchmark reports a variety of data for the benchmark, including the number of threads, array size, etc. For each run, all of the data generated should be included on the benchmark submission DVD as well as the compilation details.


Dr. John McCalpin is the developer and maintainer of the STREAM benchmark.