> -----Original Message----- > From: Jerin Jacob <jerinjac...@gmail.com> > Sent: Tuesday, September 27, 2022 6:56 PM > To: Gujjar, Abhinandan S <abhinandan.guj...@intel.com> > Cc: Volodymyr Fialko <vfia...@marvell.com>; dev@dpdk.org; Jerin Jacob > Kollanukkaran <jer...@marvell.com>; Pavan Nikhilesh Bhagavatula > <pbhagavat...@marvell.com>; Shijith Thotton <sthot...@marvell.com>; > Hemant Agrawal <hemant.agra...@nxp.com>; Sachin Saxena > <sachin.sax...@oss.nxp.com>; Jayatheerthan, Jay > <jay.jayatheert...@intel.com>; Akhil Goyal <gak...@marvell.com>; Anoob > Joseph <ano...@marvell.com> > Subject: Re: [PATCH 1/3] eventdev: introduce event cryptodev vector type > > > > > > > > > > > > Right now we are targeting crypto_cn10k PMD and ipsec-secgw > > > > > event mode to support vectorization. > > > > Is there a way to test this? When can be dataplane changes expected? > > > > > > > If the spec looks okay, support in s/w crypto adapter and other h/w > > > PMDs can be added by respective maintainers. Currently, we are > > > adding library change, support for one PMD and an application to > > > test the feature. Feature is exposed with capability flag to not break > existing functionality. > > Got it. How do I test this feature without data plane changes? > > > Hi @Gujjar, Abhinandan S > > > If there is a way to test this, please let me know. > > Dataplane changes can be tested on the cn10k platform. > This feature is a hardware assisted feature. > > > This design is right now tested for cn10k, I am not sure this works for sw > adapter. > > SW driver support is not added in this series as in order to accept a API > change, > one would need, 1)API spec 2)One of the driver 3)Test application to exercise > the API. > > It is a similar case for all ethdev, rte_flow features etc. > Community can add SW driver support just like any other subsystem APIs. > > Also, The proposed library changes don't differentiate between SW & HW > PMDs. > The proposed changes are exposed with a capability flag and so SW crypto > adapter will not have any perf impact. > > > I need to have perf data with and without vectorization support to approve. > > On the cn10k platform, we see nearly 2.5x performance with vectorization. Eth > rx adapter already supports vectorization and this spec change is in line with > that. > > Also IPsec gateway update to exercise these APIs. See > http://patches.dpdk.org/project/dpdk/patch/20220804103626.102688-6- > vfia...@marvell.com/ > > Command to test on drivers which have this functionality. > > ./dpdk-ipsec-secgw -c 0xff0000 -a 0002:01:00.1 -a 0002:20:00.1 -a > 0002:1e:00.0 -- -P -p 0x1 -P --transfer-mode event -l --event-schedule-type > parallel --desc-nb 8192 --event-vector -f simple.conf > > sample.conf > > sp ipv4 out esp protect 19 pri 1 dst 192.18.0.0/32 sport 0:65535 dport 0:65535 > sa out 19 aead_algo aes-128-gcm aead_key > 73:69:78:74:65:65:6e:62:79:74:65:73:20:6b:65:79:64:70:64:6b mode ipv4- > tunnel src 2.1.1.1 dst 1.1.1.1 type lookaside-protocol-offload port_id 0 > > neigh port 0 d0:37:45:02:b0:d3 > rt ipv4 dst 1.1.0.0/16 port 0 > > In order to make forward progress and merge patch in RC1, I would request > 1)Review the API specific patch(eventdev: introduce event cryptodev vector > type), If spec needs to be changed to adapt any other driver(SW or HW) then > the author should address that. > 2)If you think, API usage is not enough with dpdk-ipsec-secgw application, I > think, author should update the test-eventdev application to support the new > mode.Which can be merged after RC1 as it is a test application change. > > Let us know what you think to make forward progress.
Hi @Jerin Jacob & @Volodymyr Fialko, Thanks for clarifying. We would like to have a call to discuss on this. Could you please setup a call & include @Jayatheerthan, Jay also into it? Thanks Abhinandan
