benchmark — runs the BRL-CAD benchmark
The benchmark suite will test the performance of a given system by iteratively rendering several well-known datasets into 512x512 images where performance metrics are documented and fairly well understood. The local machine performance is compared to the base system (called VGR) and a numeric "VGR" multiplier of performance is computed. This number is a simplified metric from which one may qualitatively compare cpu and cache performance, versions of BRL-CAD, and different compiler characteristics.
Run without any arguments or variables set, the benchmark suite will run a series of tests where it renders several image frames. By default, the benchmark suite will attempt to calibrate the intensity of the test frames with the performance characteristics of the machine running the tests.
There are several OPTION environment variables that will modify how the BRL-CAD benchmark behaves so that it may be run in a stand-alone environment:
RT - the ray-trace binary (e.g. ../src/rt/rt or /usr/brlcad/bin/rt) DB - the directory containing the reference geometry (e.g. ../db) PIX - the directory containing the reference images (e.g. ../pix) LOG - the directory containing the reference logs (e.g. ../pix) CMP - the name of a pixel comparison tool (e.g. ./pixcmp or cmp) ELP - the name of an elapsed time tool (e.g. ../sh/elapsed.sh) TIMEFRAME - minimum number of seconds each ray-trace frame should take MAXTIME - maximum number of seconds to spend on any test DEVIATION - minimum sufficient % deviation from the average AVERAGE - how many frames to average together VERBOSE - turn on extra debug output (set to yes) QUIET - turn off all output (set to yes)
The TIMEFRAME, MAXTIME, DEVIATION, and AVERAGE options control how the benchmark will proceed including how long it should take. Each individual benchmark run will consume at least a minimum TIMEFRAME of wallclock time so that the results can stabilize. After consuming at least the minimum TIMEFRAME, additional frames may be computed until the standard deviation from the last AVERAGE count of frames is below the specified DEVIATION. When a test is run and it completes in less than TIMEFRAME, the raytrace is restarted using double the number of rays from the previous run. If the machine is fast enough, the benchmark may accelerate the number or rays being fired. These additional rays are hypersampled but without any jitter, so it is effectively performing a multiplier amount of work over the initial frame.
There are various commands that may be give to the BRL-CAD benchmark that cause it to perform various actions such as invoking the computation tests or filesystem clean-up:
clean - remove the test-specific pix and log files clobber - same as clean, also benchmark.log files and user is prompted help - displays a brief usage statement instructions - displays more detailed usage instructions quiet - quell printing output (still logs) verbose - remove the test-specific pix and log files clean - remove the test-specific pix and log files run - initiate the benchmark analysis
When the benchmark completes, output should be saved to several log files including a 'summary' file containing tabulated results, a 'benchmark.log' file containing the output from a given run, and multiple log files for each test frame. Use the clean and clobber commands during execution to remove the files generated during the benchmark.
The clean command removes the test pix and log files. The clobber command removes those same files as well as any *benchmark.log files encountered, and prompting the user beforehand. The generated tabular summary file will not be removed automatically regardless of invocation command. The summary file must always be manually deleted.
Please send your BRL-CAD Benchmark results to the developers along with
detailed system information to <devs@brlcad.org> Include at least:
0) Compiler name and version (e.g. gcc --version) 1) CPU configuration (e.g. cat /proc/cpuinfo or hinv or sysctl -a) 2) Cache (data and/or instruction) details for L1/L2/L3 and system (e.g. cat /proc/cpuinfo or hinv or sysctl -a) 3) All generated log files (i.e. *.log* after benchmark completes) 4) Anything else you think might be relevant to performance
benchmark run default run of the suite, taking approximately 10 minutes
benchmark run TIMEFRAME=1 quick test run for testing functionality and performance ballpark
benchmark run DEVIATION=1 TIMEFRAME=60 MAXTIME=600 excessive analysis, attempt to stabilize within 1 percent deviation with each frame taking at least 60 seconds but no more than 10 minutes per test (the entire analysis will probably take 30 to 60 minutes) benchmark run QUIET=1 -P1 perform a benchmark analysis only using one CPU and only logging results to a file
benchmark clean delete all of the log and pix image files generated during a benchmark analysis, leaving only the summary file and any *benchmark.log files