summaryrefslogtreecommitdiffstats
path: root/src/dmclock/benchmark/README.md
diff options
context:
space:
mode:
Diffstat (limited to 'src/dmclock/benchmark/README.md')
-rw-r--r--src/dmclock/benchmark/README.md42
1 files changed, 42 insertions, 0 deletions
diff --git a/src/dmclock/benchmark/README.md b/src/dmclock/benchmark/README.md
new file mode 100644
index 000000000..d945e986f
--- /dev/null
+++ b/src/dmclock/benchmark/README.md
@@ -0,0 +1,42 @@
+# dmclock benchmarking
+
+**IMPORTANT**: now that K_WAY_HEAP is no longer allowed to have the
+value 1, the shell and Python scripts that generate the PDFs no longer
+work exactly correctly. Some effort to debug is necessary.
+
+This directory contains scripts to evaluate effects of different
+branching-factors (k=1 to k=11) in the IndirectIntrusiveHeap
+data-structure. IndirectIntrusiveHeap is now a k-way heap, so finding
+an ideal value for k (i.e., k=2 or k=3) for a particular work-load is
+important. Also, it is well-documented that the right choice of
+k-value improves the caching behaviour [Syed -- citation needed
+here]. As a result, the overall performance of an application using
+k-way heap increases significantly [Syed -- citation needed here].
+
+A rule of thumb is the following:
+ if number of elements are <= 6, use k=1
+ otherwise, use k=3.
+
+## Prerequisites
+
+requires python 2.7, gnuplot, and awk.
+
+## Running benchmark
+
+./run.sh [name_of_the_output] [k_way] [repeat] # [Syed -- last two command line args do not work]
+
+The "run.sh" script looks for config files in the "configs" directory,
+and the final output is generated as
+"name_of_the_output.pdf". Internally, "run.sh" calls other scripts
+such as data_gen.sh, data_parser.py, and plot_gen.sh.
+
+## Modifying parameters
+
+To modify k-value and/or the amount of times each simulation is
+repeated, modify the following two variables in "run.sh" file:
+
+ k_way=[your_value]
+ repeat=[your_value]
+
+For example, k_way=3 means, the benchmark will compare simulations
+using 1-way, 2-way, and 3-way heaps.