Index of /anton/bplat

[ICO]NameLast modifiedSizeDescription

[DIR]Parent Directory  -  
[   ]Results11-Jun-2022 14:32 5.5K 
[   ]bplat-0.1.tar.gz24-Jul-2004 22:24 3.7K 

This is bplat, a memory latency benchmark that tries to improve on
lmbench's lat_mem_rd by not being predictable by stride predictors.

This benchmark is based on a program posted by Bernd Paysan in
comp.arch (see the file Postings for some postings about that), and I
just modified it and packaged it up for easy benchmarking.  I am not
sure that I got all the parameters right (especially the cs and cl
parameters to latency() in bplat.c), so take this as a first attempt.
Hopefully Bernd will pick it up and correct it if it is wrong.


Just type:

tar xvfz bplat-$VERSION.tar.gz
cd bplat-$VERSION

and you will get a list of working set sizes and the load-to-load
latency for these working sets, e.g.:

262144   8.8

which means that for a working set of 256KB, the program measured
8.8ns load-to-load latency.

You can adjust the MSPERTICK value if your OS has longer ticks than
10ms; e.g. with a 60HZ clock, you can use


You can specify the working set sizes to measure through SIZES, e.g.:

make SIZES="33554432 67108864"

Note that apparently only powers of 2 work correctly at the moment.


you can get bplat at


This program is free software under BSD or MIT/X11 license.

- anton
M. Anton Ertl                    Some things have to be seen to be believed Most things have to be believed to be seen