Awesome

FIRESTARTER - A Processor Stress Test Utility

FIRESTARTER maximizes the energy consumption of 64-Bit x86 processors by generating heavy load on the execution units as well as transferring data between the cores and multiple levels of the memory hierarchy.

WARNING: This software REALLY comes WITHOUT ANY WARRANTY! Some systems cannot handle the high load and crash due to overheating. It cannot be ruled out that the stress test damages the hardware! USE AT YOUR OWN RISK!

Supported CPU Microarchitectures

Intel Nehalem, Westmere, Sandy Bridge, Ivy Bridge, Haswell, Skylake, Knights Landing
AMD Bulldozer (experimental), Zen, Zen+, Zen2

Usage and Options

Usage:
  ./FIRESTARTER [OPTION...]

Information Options:
  -h, --help [=SECTION(=)]      Display usage information. SECTION can be any of:
                                information | general | specialized-workloads | debug
                                | measurement | optimization
  -v, --version                 Display version information
  -c, --copyright               Display copyright information
  -w, --warranty                Display warranty information
  -q, --quiet                   Set log level to Warning
  -r, --report                  Display additional information (overridden by -q)
      --debug                   Print debug output
  -a, --avail                   List available functions

General Options:
  -i, --function ID             Specify integer ID of the load-function to be
                                used (as listed by --avail)
  -f, --usegpufloat             Use single precision matrix multiplications
                                instead of default
  -d, --usegpudouble            Use double precision matrix multiplications
                                instead of default
  -g, --gpus arg                Number of gpus to use, default: -1 (all)
  -m, --matrixsize arg          Size of the matrix to calculate, default: 0 (maximum)
  -t, --timeout TIMEOUT         Set the timeout (seconds) after which FIRESTARTER
                                terminates itself, default: 0 (no timeout)
  -l, --load LOAD               Set the percentage of high CPU load to LOAD
                                (%) default: 100, valid values: 0 <= LOAD <=
                                100, threads will be idle in the remaining time,
                                frequency of load changes is determined by -p. This option
                                does NOT influence the GPU
                                workload!
  -p, --period PERIOD           Set the interval length for CPUs to PERIOD
                                (usec), default: 100000, each interval contains
                                a high load and an idle phase, the percentage
                                of high load is defined by -l.
  -n, --threads COUNT           Specify the number of threads. Cannot be
                                combined with -b | --bind, which impicitly
                                specifies the number of threads.
  -b, --bind CPULIST            Select certain CPUs. CPULIST format: "x,y,z",
                                "x-y", "x-y/step", and any combination of the
                                above. Cannot be combined with -n | --threads.
      --error-detection         Enable error detection. This aborts execution when the
                                calculated data is corruped by errors. FIRESTARTER must run
                                with 2 or more threads for this feature. Cannot be used with
                                -l | --load and --optimize.

Specialized workloads:
      --list-instruction-groups
                                List the available instruction groups for the
                                payload of the current platform.
      --run-instruction-groups GROUPS
                                Run the payload with the specified
                                instruction groups. GROUPS format: multiple INST:VAL
                                pairs comma-seperated.
      --set-line-count arg      Set the number of lines for a payload.

Debugging:
      --allow-unavailable-payload

      --dump-registers [=DELAY(=10)]
                                Dump the working registers on the first
                                thread. Depending on the payload these are mm, xmm,
                                ymm or zmm. Only use it without a timeout and
                                100 percent load. DELAY between dumps in secs.
      --dump-registers-outpath PATH
                                Path for the dump of the output files. If
                                PATH is not given, current working directory will
                                be used.
Measurement:
      --list-metrics            List the available metrics.
      --metric-from-stdin NAME  Add a metric NAME with values from stdin.
                                Format of input: "NAME TIME_SINCE_EPOCH VALUE\n".
                                TIME_SINCE_EPOCH is a int64 in nanoseconds. VALUE is a
                                double. (Do not forget to flush
                                lines!)
      --measurement             Start a measurement for the time specified by
                                -t | --timeout. (The timeout must be greater
                                than the start and stop deltas.) Cannot be
                                combined with --optimize.
      --measurement-interval arg
                                Interval of measurements in milliseconds, default: 100
      --start-delta N           Cut of first N milliseconds of measurement, default: 5000
      --stop-delta N            Cut of last N milliseconds of measurement, default: 2000
      --preheat N               Preheat for N seconds, default: 240

Optimization:
      --optimize arg            Run the optimization with one of these algorithms: NSGA2.
                                Cannot be combined with --measurement.
      --optimize-outfile arg    Dump the output of the optimization into this
                                file, default: $PWD/$HOSTNAME_$DATE.json
      --optimization-metric arg
                                Use a metric for optimization. Metrics listed
                                with cli argument --list-metrics or specified
                                with --metric-from-stdin are valid.
      --individuals arg         Number of individuals for the population. For
                                NSGA2 specify at least 5 and a multiple of 4,
                                default: 20
      --generations arg         Number of generations, default: 20
      --nsga2-cr arg            Crossover probability. Must be in range [0,1[
                                default: 0.6
      --nsga2-m arg             Mutation probability. Must be in range [0,1]
                                default: 0.4

Examples:
  ./FIRESTARTER                 starts FIRESTARTER without timeout
  ./FIRESTARTER -t 300          starts a 5 minute run of FIRESTARTER
  ./FIRESTARTER -l 50 -t 600    starts a 10 minute run of FIRESTARTER with
                                50% high load on CPUs and full load on GPUs
  ./FIRESTARTER -l 75 -p 20000000
                                starts FIRESTARTER with an interval length
                                of 2 sec, 1.5s high load and 0.5s idle
                                on CPUs and full load on GPUs
  ./FIRESTARTER --measurement --start-delta=300000 -t 900
                                starts FIRESTARTER measuring all available
                                metrics for 15 minutes disregarding the first
                                5 minutes and last two seconds (default to `--stop-delta`)
  ./FIRESTARTER -t 20 --optimize=NSGA2 --optimization-metric sysfs-powercap-rapl,perf-ipc
                                starts FIRESTARTER optimizing with the sysfs-powercap-rapl
                                and perf-ipc metric. The duration is 20s long. The default
                                instruction groups for the current platform will be used.

Building FIRESTARTER

FIRESTARTER can be build under Linux, Windows and macOS with CMake.

GCC (>=7) or Clang (>=9) is supported.

CMake option	Description
`FIRESTARTER_BUILD_TYPE`	Can be any of `FIRESTARTER`, `FIRESTARTER_CUDA`, `FIRESTARTER_ONEAPI`, or `FIRESTARTER_HIP`. Default `FIRESTARTER`
`FIRESTARTER_LINK_STATIC`	Link FIRESTARTER as a static binary. Note, dlopen is not supported in static binaries. This option is not available on macOS or with CUDA or OneAPI enabled. Default `ON`
`FIRESTARTER_BUILD_HWLOC`	Build hwloc dependency. Default `ON`
`FIRESTARTER_THREAD_AFFINITY`	Enable FIRESTARTER to set affinity to hardware threads. Default `ON`

When building FIRESTARTER_ONEAPI make sure that the Intel Math Kernel Library (MKL) and the complier icx can be found. These will be used to build FIRESTARTER, while dependencies will be build with $CC and $CXX respectively.

Metrics

The Linux version of FIRESTARTER supports to collect metrics during runtime. Available metrics can be shown with --list-metrics. Default metrics are perf-ipc, perf-freq, ipc-estimate and sysfs-powercap-rapl.

Custom Metrics

If one would like to use custom metrics, e.g. an external power measurement, --metric-from-stdin=NAME allows metric values to be passed via stdin in the following format: {NAME} {TIME SINCE EPOCH IN NS} {ABSOLUT METRIC VALUE (double)}\n. See here for a basic example.

Metric Recording

The Linux version of FIRESTARTER has the option to output the collected metric values by specifying --measurement. Options --start-delta (default 5000ms) and --stop-delta (default 2000ms) specify a time in milliseconds in which metric values should be ignored. After a run, the output will be printed in CSV format to stdout.

Measurement Example

Measure all available metrics for 15 minutes disregarding the first 5 minutes and last two seconds (default to --stop-delta).

FIRESTARTER --measurement --start-delta=300000 -t 900

Optimization

The Linux version of FIRESTARTER has the option to optimize itself using evolutionary algorithms. It currently supports the multiobjective algorithm NSGA2, selected by --optimize=NSGA2.

The evolutionary algorithm evaluates individuals one after another. Each evaluation of a given individual is -t | --timeout seconds long. Selecting a long enough time for letting the power consumption stabilize, but not too long as this will leed to a much longer optimization timespan. During this time metrics are collected and the selected metrics (--optimization-metrics) are used for assigning a fitness (specify at least two metrics for a multiobjective algorithm). The optimization result depends on the accuracy of the power measurment. The sysfs-powercap-rapl metric correlates strongly with the actual system power consumption on Intel processors since Haswell, see An Energy Efficiency Feature Survey of the Intel Haswell Processor.

Individuals are made of different instruction groups and their ratios to one another. Without specifying the --run-instruction-groups option, preselected instruction groups will be used for optimization. Setting this option allows the user to select different instruction group, e.g. for optimizing FIRESTARTER on a not yet optimized microarchitecture. The format of this is the same as shown by -a | --avail. All available instruction groups can be listed with --list-instruction-groups.

The number of individuals per generation (--individuals) and the number of generations (--generation) are both set 20 per default.

Before the optimization runs, a user-defined period of preheating of the CPU is carried out. Option --preheat defaults to 240 seconds.

After the optimization finishes the acquired data will be written to {HOSTNAME}_${STARTTIME}.json if not specified otherwise with the option --optimize-outfile. An IPython Notebook is provided for basic visualization.

The NSGA2 Algorithm

The NSGA2 algorithm, as described in A fast and elitist multiobjective genetic algorithm: NSGA-II, is a multiobjective algorithm allowing FIRESTARTER to optimize with two (or more) metrics. This is relevant because adding the IPC (instruction per cycle) metric supports the optimization algorithm to converge towards higher power consumption. Parameters of the algorithm can be tweaked using --nsga2-cr and --nsga2-m.

Optimization Examples

Optimize FIRESTARTER with NSGA2 and sysfs-powercap-rapl and perf-ipc metric. The duration for the evaluation of a setting is 20s. -a | --avail lists the default instruction groups for the current platform.

FIRESTARTER -t 20 --optimize=NSGA2 --optimization-metric sysfs-powercap-rapl,perf-ipc

If perf-ipc is not available use ipc-estimate

FIRESTARTER -t 20 --optimize=NSGA2 --optimization-metric sysfs-powercap-rapl,ipc-estimate

OneAPI

FIRESTARTER_ONEAPI needs to find certain libraries, which are installed with the Intel OneAPI toolkits, since some of these libraries cannot be linked statically. Make sure that they are installed and $LD_LIBRARY_PATH is set-up correctly. Make also sure that you have the correct drivers for your GPU installed so that OneAPI can apply work to it.

Reference

A detailed description can be found in the following paper. Please cite this if you use FIRESTARTER for scientific work.

Daniel Hackenberg, Roland Oldenburg, Daniel Molka, and Robert Schöne Introducing FIRESTARTER: A processor stress test utility (IGCC 2013)

Additional information: https://tu-dresden.de/zih/forschung/projekte/firestarter.

Contact

Daniel Hackenberg < daniel dot hackenberg at tu-dresden.de >

License

GNU GENERAL PUBLIC LICENSE Version 3

FIRESTARTER - A Processor Stress Test Utility Copyright (C) TU Dresden, Center for Information Services and High Performance Computing

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.

Licenses of Used Software Packages

This program contains a slightly modified version of the implementation of the NSGA2 algorithm from esa/pagmo2 licensed under LGPL or GPL v3.

This program incorporates following libraries asmjit/asmjit licensed under zlib, open-mpi/hwloc licensed under BSD 3-clause, jarro2783/cxxopts licensed under MIT, nlohmann/json licensed under MIT and tud-zih-energy/nitro licensed under BSD 3-clause.