Re: i386 and amd64 testbeds now use NVMM
Jukka Ruohonen wrote: > > This reduces the time it takes to run the test suite from more than > > 20 hours to about 3-4 hours. Many thanks to Maxime Villard for making > > this possible by writing NVMM. > > Does this mean that the amount of test runs increases accordingly (i.e., > to about six runs per 24h)? It's a bit more complicated than that. Since multiple source versions are tested in parallel, the i386 tests have been achieving a throughput of more than six runs per 24 h even before the switch to NVMM. Using NVMM frees up a significant amount of CPU, but the builds and the sparc tests still use as much CPU as before. So the overall throughput of the server has increased, but by a smaller factor than the latency of the i386 and amd64 tests. -- Andreas Gustafsson, g...@gson.org
Re: i386 and amd64 testbeds now use NVMM
On Mon, Aug 10, 2020 at 09:19:00PM +0300, Andreas Gustafsson wrote: > This reduces the time it takes to run the test suite from more than > 20 hours to about 3-4 hours. Many thanks to Maxime Villard for making > this possible by writing NVMM. Does this mean that the amount of test runs increases accordingly (i.e., to about six runs per 24h)? - Jukka
i386 and amd64 testbeds now use NVMM
Hi all, The TNF testbed is now using NVMM for the i386 and amd64 tests: http://releng.netbsd.org/b5reports/i386/ http://releng.netbsd.org/b5reports/amd64/ This reduces the time it takes to run the test suite from more than 20 hours to about 3-4 hours. Many thanks to Maxime Villard for making this possible by writing NVMM. The switch to NVMM was made yesterday, but since the testbed may test source versions out of order, there is not necessarily an unambiguous transition point in terms of -current source dates. To determine whether a given test run was made using NVMM or not, look for "-accel nvmm" in the qemu command line in the console log. Some test cases that were previously failing are now passing and vice versa. For example, kernel/t_trapsignal:fpe_* now pass, but lib/libpthread/t_condwait:* now fail (these contain a work-around for the qemu timing issues of PR 43997, but now fail to detect that they are running under qemu). -- Andreas Gustafsson, g...@gson.org