Commit ec8ab103 authored by Jan Eitzinger's avatar Jan Eitzinger
Browse files

Merge branch 'master' of

parents 987fdbf9 de0e6153
......@@ -88,9 +88,9 @@ To run the benchmark call:
The benchmark will output the results similar to the stream benchmark. Results are validated.
For threaded execution it is recommended to control thread affinity.
We recommend to use likwid-pin for benchmarking:
We recommend to use likwid-pin for setting the number of threads used and to control thread affinity:
likwid-pin -c 0-3 ./bwbench-GCC
likwid-pin -C 0-3 ./bwbench-GCC
Example output for threaded execution:
......@@ -118,3 +118,42 @@ SDaxpy: 46822.63 23411.32 0.0281 0.0273 0.0325
Solution Validates
## Scaling runs
Apart from the highest sustained memory bandwidth also the scaling behavior within memory domains is a important system property.
There is a helper script included in util (``````) that creates a text result file from multiple runs that can be used as input to plotting applications as gnuplot and xmgrace.
This involves two steps: Executing the benchmark runs and creating the data file.
To run the benchmark for different thread counts within a memory domain execute (this assumes bash or zsh):
$ for nt in 1 2 4 6 8 10; do likwid-pin -q -C E:M0:$nt:1:2 ./bwbench-ICC > dat/emmy-$nt.txt; done
It is recommended to just use one thread per core in case the processor supports hyperthreading.
Use whatever stepping you like, here a stepping of two was used.
The ```-q``` option suppresses output from ```likwid-pin```.
Above line uses the expression based syntax, on systems with hyperthreading enabled (check with, e.g., ```likwid-topology```) you have to skip the other hardware threads on each core.
For above system with 2 hardware threads per core this results in ```-C E:M0:$nt:1:2```, on a system with 4 hardware threads per core you would need ```-C E:M0:$nt:1:4```.
The string before the dash (here emmy) can be arbitrary, but the the extraction script expects the thread count after the dash.
Also the file ending has to be ```.txt```.
Please check with a text editor on some result files if everything worked as expected.
To extract the results and output in a plottable format execute:
./ ./dat
The script will pick up all result files in the directory specified and create a column format output file.
In this case:
#nt Init Sum Copy Update Triad Daxpy STriad SDaxpy
1 4109 11900 5637 8025 7407 9874 8981 11288
2 8057 22696 11011 15174 14821 18786 17599 21475
4 15602 39327 21020 28197 27287 33633 31939 37146
6 22592 45877 29618 37155 36664 40259 39911 41546
8 28641 46878 35763 40111 40106 41293 41022 41950
10 33151 46741 38187 40269 39960 40922 40567 41606
Please be aware the the single core memory bandwidth as well as the scaling behavior depends on the frequency settings.
......@@ -4,7 +4,9 @@ bwBench.c contains a single file version of The Bandwidth Benchmark that is tail
It should compile with any C99 compiler.
# Benchmarking skript
# Benchmarking skripts
## to determine the absolute highest main memory bandwidth
A wrapper scripts in perl ( and python ( are also provided to scan ranges of thread counts and determine the absolute highest sustained main memory bandwidth. In order to use it `likwid-pin` has to be in your path. The script has three required and one optional command line arguments:
......@@ -18,3 +20,7 @@ The script will always use physical cores only, where two SMT threads is the def
$./ ./bwbench-GCC 14-24 10 1
## to generate a plottable output files from multiple scaling runs
Please see how to use it in the toplevel [README](
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment