Hello everyone, With the collaboration of Petr Holasek we released a first 0.1 version of the AutoNUMA benchmark. It's now trivial to run it without the chance of mistakes, and you can also see how fast the NUMA algorithms in the kernel converge the load by checking the pdf charts it creates after each benchmark completes. This benchmark can also setup hard/inverse bindings to benchmark the hardware and measure the best/worst case (when run with hard bindings no memory migration will ever happen and it starts computing in the ideal memory layout, so it'll always be slightly faster than the non-hard binding case). To run it you just need numactl, gnuplot and gcc installed. After cloning this git repo: git clone git://gitorious.org/autonuma-benchmark/autonuma-benchmark.git you can simply run it as root (or with sudo prefix): ./start_bench.sh -A The above command will run all tests including the hard/inverse binds (which require root privileges). The first objective of this benchmark is to be able to track regression in AutoNUMA, but it's also useful to to compare the results of this benchmark on upstream, tip.git, and aa.git with AutoNUMA enabled. If you're only going to compare different NUMA placement algorithms in different kernels, you can skip the inverse/hard bind tests to speed up the benchmarking effort (the hard/invers bind tests should result always the same). To skip the hard/inverse bind tests you can run it like this: ./start_bench.sh -s -t For a basic and quick benchmark, you can run it without parameters: ./start_bench.sh If you encounter any problem or if you're interested to contribute please CC Petr too. Thanks, Andrea -- To unsubscribe from this list: send the line "unsubscribe linux-numa" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
