Message ID | cover.1655219150.git.asml.silence@gmail.com (mailing list archive) |
---|---|
Headers | show |
Series | single-issuer and poll benchmark | expand |
On 6/14/22 9:27 AM, Pavel Begunkov wrote: > Add some tests to check the kernel enforcing single-issuer right, and > add a simple poll benchmark, might be useful if we do poll changes. Should we add a benchmark/ or something directory rather than use examples/ ? I know Dylan was looking at that at one point. I don't feel too strongly, as long as it doesn't go into test/.
On 6/14/22 16:30, Jens Axboe wrote: > On 6/14/22 9:27 AM, Pavel Begunkov wrote: >> Add some tests to check the kernel enforcing single-issuer right, and >> add a simple poll benchmark, might be useful if we do poll changes. > > Should we add a benchmark/ or something directory rather than use > examples/ ? > > I know Dylan was looking at that at one point. I don't feel too > strongly, as long as it doesn't go into test/. I don't care much myself, I can respin it once (if) the kernel side is queued.
On 6/14/22 9:48 AM, Pavel Begunkov wrote: > On 6/14/22 16:30, Jens Axboe wrote: >> On 6/14/22 9:27 AM, Pavel Begunkov wrote: >>> Add some tests to check the kernel enforcing single-issuer right, and >>> add a simple poll benchmark, might be useful if we do poll changes. >> >> Should we add a benchmark/ or something directory rather than use >> examples/ ? >> >> I know Dylan was looking at that at one point. I don't feel too >> strongly, as long as it doesn't go into test/. > > I don't care much myself, I can respin it once (if) the kernel > side is queued. I'm leaning towards just using examples/ - but maybe Dylan had some reasoning for the new directory. CC'ed.
On Tue, 2022-06-14 at 09:51 -0600, Jens Axboe wrote: > On 6/14/22 9:48 AM, Pavel Begunkov wrote: > > On 6/14/22 16:30, Jens Axboe wrote: > > > On 6/14/22 9:27 AM, Pavel Begunkov wrote: > > > > Add some tests to check the kernel enforcing single-issuer > > > > right, and > > > > add a simple poll benchmark, might be useful if we do poll > > > > changes. > > > > > > Should we add a benchmark/ or something directory rather than use > > > examples/ ? > > > > > > I know Dylan was looking at that at one point. I don't feel too > > > strongly, as long as it doesn't go into test/. > > > > I don't care much myself, I can respin it once (if) the kernel > > side is queued. > > I'm leaning towards just using examples/ - but maybe Dylan had some > reasoning for the new directory. CC'ed. > I wanted to have some common framework for benchmarks to make it easy to write new ones and get some nice numbers. In that case it would make sense for them to be in one place. But I haven't finished it yet, so probably for now examples/ is best and if/when I finish it I can port these over.
On 6/15/22 02:23, Dylan Yudaken wrote: > On Tue, 2022-06-14 at 09:51 -0600, Jens Axboe wrote: >> On 6/14/22 9:48 AM, Pavel Begunkov wrote: >>> On 6/14/22 16:30, Jens Axboe wrote: >>>> On 6/14/22 9:27 AM, Pavel Begunkov wrote: >>>>> Add some tests to check the kernel enforcing single-issuer >>>>> right, and >>>>> add a simple poll benchmark, might be useful if we do poll >>>>> changes. >>>> >>>> Should we add a benchmark/ or something directory rather than use >>>> examples/ ? >>>> >>>> I know Dylan was looking at that at one point. I don't feel too >>>> strongly, as long as it doesn't go into test/. >>> >>> I don't care much myself, I can respin it once (if) the kernel >>> side is queued. >> >> I'm leaning towards just using examples/ - but maybe Dylan had some >> reasoning for the new directory. CC'ed. >> > > I wanted to have some common framework for benchmarks to make it easy > to write new ones and get some nice numbers. In that case it would make > sense for them to be in one place. This sounds great, it's really nice to have a series of standard performance tests rather than temporary tests created along with new kernel patches. A framework is something nicer. Actually there are people doing this though just a simple one, FWIW: https://gitee.com/anolis/perf-test-for-io_uring > > But I haven't finished it yet, so probably for now examples/ is best > and if/when I finish it I can port these over. >