Hi Bhanu, Perhaps we can aim at fine-tuning the layout of performance-critical structs once at the end of a release cycle when all development is completed. That way we don't hinder productivity during the release cycle and still won't unnecessarily loose performance.
And we should definitely apply the now corrected xmalloc_cacheline() for DPDK datapath structs of size 64B or larger to have a well-defined cache alignment base. BR, Jan > -----Original Message----- > From: Bodireddy, Bhanuprakash [mailto:[email protected]] > Sent: Thursday, 07 December, 2017 21:22 > To: [email protected]; Ilya Maximets <[email protected]>; > [email protected] > Cc: Heetae Ahn <[email protected]>; Fischetti, Antonio > <[email protected]>; Loftus, Ciara <[email protected]>; > Kevin Traynor <[email protected]>; Jan Scheurich > <[email protected]>; Stokes, Ian <[email protected]> > Subject: RE: [PATCH v6 1/7] dpif-netdev: Refactor PMD thread structure for > further extension. > > >On 07/12/17 14:28, Ilya Maximets wrote: > >> Thanks for review, comments inline. > >> > >> On 07.12.2017 15:49, Eelco Chaudron wrote: > >>> On 01/12/17 16:44, Ilya Maximets wrote: > >>>> This is preparation for 'struct dp_netdev_pmd_thread' modification > >>>> in upcoming commits. Needed to avoid reordering and regrouping while > >>>> replacing old and adding new members. > >>>> > >>> Should this be part of the TX batching set? Anyway, I'm ok if it's > >>> not stalling the approval :) > >> Unfortunately yes, because members reordered and regrouped just to > >> include new members: pmd->ctx and pmd->n_output_batches. This could > >> not be a standalone change because adding of different members will > >> require different regrouping/ reordering. I moved this change to a > >> separate patch to not do this twice while adding each member in patches > >2/7 and 6/7. > >> > >> Anyway, as I mentioned in cover letter, I still prefer reverting of > >> the padding at all by this patch: > >> > >> https://mail.openvswitch.org/pipermail/ovs-dev/2017-November/341153.html > > I understand that with PADDED_MEMBERS macro it was slightly tricky to extend > or reorganize the structure and so suggested 'pahole'. > But I see that the problem hasn't gone and still there are some strong > opinions on reverting the earlier effort. > > I don’t mind reverting the patch but would be nice if the changes to this > structure are made keeping alignment in mind. > > - Bhanuprakash. _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
