On Thu, Jan 12, 2017 at 6:27 AM, Savolainen, Petri (Nokia - FI/Espoo) <petri.savolai...@nokia-bell-labs.com> wrote: > > >> -----Original Message----- >> From: Bill Fischofer [mailto:bill.fischo...@linaro.org] >> Sent: Thursday, January 12, 2017 2:22 PM >> To: Savolainen, Petri (Nokia - FI/Espoo) <petri.savolainen@nokia-bell- >> labs.com> >> Cc: lng-odp@lists.linaro.org >> Subject: Re: [lng-odp] [API-NEXT PATCHv6 2/5] linux-generic: packet: >> implement reference apis >> >> On Thu, Jan 12, 2017 at 6:12 AM, Savolainen, Petri (Nokia - FI/Espoo) >> <petri.savolai...@nokia-bell-labs.com> wrote: >> > This patch is now merged, although I had some doubts that it has bad >> impact on performance. Here are some performance results for couple of >> simple, single thread packet alloc/free test cases. No references >> involved, just plain packets as before. >> > >> > Test results before and after "linux-generic: packet: implement >> reference apis": >> > >> > CPU cycles per operation >> > before after >> > packet_alloc 127.8 250.9 +96 % >> > packet_alloc_multi 873.7 1538.8 +76 % >> > packet_free 31 116.8 +277 % >> > packet_free_multi 214.5 1369.2 +538 % >> > packet_alloc_free 73.4 193.7 +164 % >> > packet_alloc_free_multi 286.1 1228.8 +330 % >> > >> > >> > Huge performance degradation. Numbers are now many times worse than >> before or after my optimizations. To me this shows that almost a complete >> rewrite (or revert) is needed. >> >> My guess is this is due to the atomics needed for reference counting >> not being properly inlined internally. I know you did similar >> optimizations for ticketlocks. Let me look into this and post a patch >> that does the same for the atomics. > > Atomics are inlined already. Also atomic operations should not be required if > there's a single refence.
The ref_count still needs to be initialized on alloc and decremented on free to see if the packet should really be freed. The only way to know whether you're dealing with a single reference or not is to check. Do you have a breakdown of where the hotspots are? These numbers do seem high. > > -Petri > >
