We have the same kind of "issue" with our port. We don't want (more precisely cannot for the moment) more than a thread per core.
The way we handled it is by tweaking the implementation of cpu_count and the default_mask to ignore the first core (which is running the main). It would make sense to have a more "standard" way of expressing this than hacking the cpu_count per platform/test. On 06/29/2015 05:59 PM, Maxim Uvarov wrote: > Radu-Andrei, > > we also needed to limit number of workers for dpdk: > https://lists.linaro.org/pipermail/lng-odp-dpdk/2015-June/000844.html > > I will send this patch to main mailing list. > > Maxim. > > On 06/29/15 17:44, Bill Fischofer wrote: >> This is timely. I'll add it to the agenda for tomorrow's ODP call. >> >> On Mon, Jun 29, 2015 at 9:43 AM, Radu-Andrei Bulie <radu.bu...@freescale.com >> <mailto:radu.bu...@freescale.com>> wrote: >> >> Hi, >> >> There are two items I would like to discuss: >> >> 1.I have sent an email regarding the schedule tests issues (this >> could be a general issue) >> >> To relieve from a search on the mailing list I will paste the >> content (and make a resumee): >> >> “I have some observations regarding the odp_scheduler test >> functionality. >> >> Scheduler validation creates two kind of tests: >> >> -single threaded tests – that uses the cunit main thread for >> initialization. >> >> -multithreaded tests – that create a number of threads equal with >> the number of cores in the system. >> >> As I said in an older post (regarding the classification tests) >> there could be a problem on some platforms when the schedule >> function is called >> >> on a core that is used by linux(that is –schedule cannot be call >> on any core). This issue could happen on both kind of tests. >> >> The same approach as in the main applications should be applied >> regarding tests (e.g core 0 is for linux and the other for odp >> threads) >> >> Another possible problem is that the tests use a *while(1)* – and >> cycle until all the expected packets are received on the scheduled >> queues. >> >> That means that the queues will be scheduled on each of the cores >> where the odp_threads were created. I think that there is no >> guarantee on this and there is a possibility >> >> that some of the threads will remain in the *while(1) –(queues >> will not be scheduled on those cores where the threads were >> running, the number of expected frames is not reached) *and thus >> the tests will hang.” >> >> Shortly said – the multithreaded tests from schedule create a >> number of worker threads equal with the number of cores in the >> system ( thus running one of the worker threads on the control >> core. As I said(in previous emails), this could be a problem – >> because the tests do not respect the model with a control core, >> and data path cores. >> >> In this case, on our platform, the multithreaded tests fail >> (they hang- because nothing will be scheduled on the control core, >> and as I said in the email >> >> there is a while loop that cycles until a number of packets is >> reached). Is there any intention to fix this? >> >> 2.We made an internal patch on the cunit common sources so that we >> are able to run individual tests from a validation suite. >> >> For example, for odp_pktio validation test – it could be run only >> – “pktio poll queues” test. Generally a user will be able to run >> one or multiple tests >> >> from a validation test (thus validating only those components in >> which he is interested). I want to know if this could be >> considered a useful modification and if the answer was >> affirmative, we would push the patch to the list. >> >> Regards, >> >> Radu >> >> >> _______________________________________________ >> lng-odp mailing list >> lng-odp@lists.linaro.org <mailto:lng-odp@lists.linaro.org> >> https://lists.linaro.org/mailman/listinfo/lng-odp >> >> >> >> >> _______________________________________________ >> lng-odp mailing list >> lng-odp@lists.linaro.org >> https://lists.linaro.org/mailman/listinfo/lng-odp > > _______________________________________________ > lng-odp mailing list > lng-odp@lists.linaro.org > https://lists.linaro.org/mailman/listinfo/lng-odp _______________________________________________ lng-odp mailing list lng-odp@lists.linaro.org https://lists.linaro.org/mailman/listinfo/lng-odp
