Re: [EXT] [vpp-dev] Marvell's AI from last call

Hi, Initially no, but later I did, and still got slightly better performance with DPDK. I totally agree that MUSDK plugin could outperform DPDK plugin, but it currently does not (or I am doing something wrong, what it very possible). And it is the MUSDK plugin which introduces additional

Re: [EXT] [vpp-dev] Marvell's AI from last call

Dear Szymon, It is reasonably to expect that copying data from NIC descriptor to VPP buffer metadata should be faster than copying NIC descriptor to DPDK metadata and then copying same data from DPDK metadata into vpp metadata. Do we agree on that? It might be that i wrote crapy code in

Re: [EXT] [vpp-dev] Marvell's AI from last call

Ok, sorry, I haven't disabled dpdk_plugin.so, with it disabled I get only slightly better performance with Marvell DPDK (200 Kpps or something like that). Regards, Szymon Śliwa 2018-05-24 12:14 GMT+02:00 Szymon Śliwa : > I observed different results - I got ~4 Mpps/core with

Re: [EXT] [vpp-dev] Marvell's AI from last call

I observed different results - I got ~4 Mpps/core with Marvell DPDK and ~3 Mpps/core with Marvell plugin, but maybe I have something misconfigured. FYI: there are problems when trying to use crypto_mrvl. Regards, Szymon Śliwa 2018-05-23 18:15 GMT+02:00 Vincent Jardin :

Re: [EXT] [vpp-dev] Marvell's AI from last call

Damjan is right. It does not make sense to over wrapup (overheads) when it is not needed.

Re: [EXT] [vpp-dev] Marvell's AI from last call

It is faster because there is no DPDK overhead — Damjan > On 23 May 2018, at 17:45, Tina Tsou wrote: > > Dear Damjan, > > Is the 30% faster because of using different applications or something else? > > > Thank you, > Tina > > On May 23, 2018, at 7:56 AM, Damjan

Re: [EXT] [vpp-dev] Marvell's AI from last call

Dear Damjan, Is the 30% faster because of using different applications or something else? Thank you, Tina On May 23, 2018, at 7:56 AM, Damjan Marion > wrote: I addef -2 on it as there is no value in having it in vpp. We already have

Re: [EXT] [vpp-dev] Marvell's AI from last call

I addef -2 on it as there is no value in having it in vpp. We already have native driver which is 30% faster. At the end dpdk driver is just a wrapper around musdk — Damjan > On 22 May 2018, at 20:05, Szymon Śliwa wrote: > > Hi, I improved the patch Natalie submitted,

Re: [EXT] [vpp-dev] Marvell's AI from last call

Hi, I improved the patch Natalie submitted, and pushed it for review on gerrit: https://gerrit.fd.io/r/#/c/12694/ . There was a small change in dpdk/device/init.c - I needed to set ignore_offload_bitfield manually, don't know what changed since I last tested it, but it did not work without it. I

Re: [EXT] [vpp-dev] Marvell's AI from last call

Fine if this is going to stay out of vpp repo, as it is duplicate functionality and also it will make dpdk plugin more messy due to external dependencies on musdk. Will be better to do that work with some NIC which have native DPDK drivers, not wrapper like in musdk case... -- Damjan > On

Re: [EXT] [vpp-dev] Marvell's AI from last call

Hi Damjan, This request was discussed in FD.io ARM call in order to optimize dpdk-input node for ARM platforms. Macchiatobin is one common platform where almost all ARM stakeholders are working on and profiling data/details can be easily shared among them. Thanks, Nitin On Wednesday 16 May

Re: [EXT] [vpp-dev] Marvell's AI from last call

Hi Natalie, Who requested that? Thanks, -- Damjan > On 16 May 2018, at 15:33, Natalie Samsonov wrote: > > Dear Damjan, > > As I understood, it was a request to use DPDK plugin with all platforms and I > was asked to upstream Marvell DPDK plugin. > > Best Regards, >