If not stated other, always the bandwidth per process is printed.

The output has the following sections:

**HEAD**-
Short information about the memory per processor,
the used cartesian dimensions,
and the used values for L (message length).
**LOOP**-
The time spent on the measurements is printed.
**LOOPLNGS**-
The automatically adapted values of
looplength are printed.
**VALUES**-
The benchmark measurements are done in this sections.
Following output is printed:
For each value of L, each method and each pattern:

- min:
- b(L) / (total number of messages of a pattern)
* (maximum of messages in one process)

i.e. the minimal bandwidth of one process (except bundary nodes in acyclic cartesian topologies).

The value is printed for each repetition. - ->
- The maximum of all repetitions and in parentheses the amount of time spent on the measurement of these values.
- avg:
- Same as
*min*, but the average of the time spent on the several processes is computed - max:
- Same as
*min*, but the maximum of the time spent on the several processes is computed

**ELAPSED**-
The time spent on the measurements and the same time computed
via the values of b(L), i.e. computed with the maximum of time
spent for each measurement on all nodes.
If the difference is more than 5% of the whole time
then this is a hint, that the benchmark was not executed on
a dedicated system or the MPI implementation has a bug.
**PATTERN**-
Description of the patterns.
**BY-METHODS**-
Separate analysis for each pattern and each communication method.
**BY-REPETITIONS**-
Separate analysis for each pattern and each repetition.
**BY-MTHD-MSGLNG**-
Separate analysis for each pattern, each communication method
and each message length.
**BY-MSGLNG**-
Separate analysis for each message length. The columns are
- The message length.
- Accumulated effective bandwidth for this message length.
- Effective bandwidth per process for this message length.
- Effective bandwidth per process, cartesian patterns only.
- Effective bandwidth per process, random patterns only.
- Effective bandwidth per process, method 1 (Sendrecv) only.
- Effective bandwidth per process, method 2 (Alltoallv) only.
- Effective bandwidth per process, method 3 (Irecv+Isend+Waitall) only.

**BY-PATTERN-MSGLNG**-
Separate analysis for each pattern and some message lengths.
The last column is the base for computing the logarithmic
average of the cartesian and random patterns.
The last line shows the effective bandwidth per process.
**BEFF**-
The effective bandwidth
**b**and the system parameters are printed._{eff} **Final result line**- Summary of the result on one line.

UP Effective Bandwidth Benchmark Pallas Effective Bandwidth Benchmark MPI at HLRS HLRS Navigation HLRS