FAQ
What optimizations does SPET use
Server Performance Evaluation Tool optimizations in a few ways:
- Sets the CPU scaling governor to
performance
. - Disables transparent hugepages.
- Disables swap.
- Before each performance test, it writes buffered metadata and data to be written to the file system. It also frees pagecache, dentries, and inodes which will return page cache / buffer cache to “free”.
- Sets the maximum call stack to unlimited.
- Sets the maximum number of file descriptors to 1048576.
- Sets the maximum number of user processes to unlimited.
- By default for compiled operations, it uses the appropriate
-march=
and-mtune=
flags for the system. Some compilations also add-O3
or-O2
if an optimization flag is not already set. - Add
fs.file-max = 1048576
to/etc/sysctl.conf
. - Add
* - nofile 1048576'
to/etc/security/limits.conf
.
Can I change the versions of packages used
Yes, you can override the package versions using by editing the
spet/package_versions.py
file. Please note that there is always a
chance that future versions are not setup the same as current versions,
so it could break some functionality in SPET.
Can I use custom C flags
Yes, you can override Server Performance Evaluation Tool’s optimizations
by editing the spet/overrides.py
file.
What math library does SPET use for LINPACK
Server Performance Evaluation Tool uses one of three math libraries for High-Performance Linpack.
- If the processor name has Intel in it, SPET will use MKL.
- If the processor name has AMD in it, SPET will use BLIS.
- All other processor names will default to use OpenBLAS.