Hi Vratko, > When I was running tests for this [0] revert, > I saw different performance consequences than expected
Hmm but [0] looks bogus to me anyway: if I'm not wrong, this prefetches the vlib_buffer pointer array instead of the vlib_buffer headers pointed to by the array. I think the correct patch should be https://gerrit.fd.io/r/c/vpp/+/35419 > What do you think, is it possible that also for VPP,> > some regressions are caused just by compiler changing the memory layout > (without any "real" change in the program logic)? This can definitely happen, I have seen this several time myself. > If yes, should we try to apply Stabilizer to VPP? I wouldn't go there: looks like an interesting but unmaintained research project. Last real commits are 9yr old and they mentioned LLVM 3.1. You're not going to compile VPP with that. If you feel like tackling it, I'd look into techniques for reproducible builds that could help by using same libraries and compilers versions, but in my experience it has been mostly frustrating to try to control what the compiler do to avoid this kind of issue. Best ben
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#20892): https://lists.fd.io/g/vpp-dev/message/20892 Mute This Topic: https://lists.fd.io/mt/89321880/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
