Re: Redirection Table

[Ivan Boule <ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>]

On 01/07/2014 02:03 PM, Stefan Baranoff wrote:
>
> All,
>
> Does this mean that an application looking at traffic in something 
> like an IP/IP or GRE tunnel with only two endpoints on the tunnels but 
> many clients behind them must do software load balancing as the 
> packets would IP only (not TCP/UDP) with the same two addresses?
>
Yes, your understanding is correct, when Intel's NICs are used. But some 
other NICs could allow deeper packet analysis.
When NICs only support IPv4/IPv6 RSS, then it is up to the network stack 
on top of the DPDK to provide the best optimal solution in software.

> How much of a penalty is there for crossing processor boundaries in 
> that case and might a 1 CPU server, while less core dense, actually 
> give better performance/watt?
>
This is hard to say. It is likely to partially depend on the 
CPU/memory/bus/ characteristics.
Note that if you consider dedicating a single core to the polling of 
packets received on all ports, then you only need a single RX queue per 
port.
By the way, in such a case, you can still configure RSS on the ports, so 
that the hardware computes the 32-bit RSS hash on each IP packet (it is 
stored into the rte_mbuf structure by the Poll Mode Driver), so that the 
RSS hash can be used in software to speed-up the assignment of IP 
packets to [a subset of] processing cores, if needed.

Regards
Ivan

> Thanks,
> Stefan
>
> Sent from my smart phone; people don't make typos, Swype does!
>
> On Jan 7, 2014 3:36 AM, "Ivan Boule" <ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org 
> <mailto:ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>> wrote:
>
>     On 01/06/2014 05:52 PM, Michael Quicquaro wrote:
>
>         Thanks for the details.  Can the hash function be modified so
>         that I can provide my own RSS function?  i.e.  my ultimate
>         goal is to provide RSS that is not dependent on packet contents.
>
>     No, the RSS function is "hard-wired" and only works on IPv4/IPv6
>     packets. All other packets are stored in the same queue (0 by
>     default).
>     You can change the RSS key used by the RSS function to compute the
>     hash value.
>     See the following testpmd command:
>
>        port config X rss-hash-key <80 hexa digits>
>
>     to set the 320-bit RSS key of port X.
>
>     Best regards,
>     Ivan
>
>         You may have seen my thread "generic load balancing".  At this
>         point, I'm realizing that the only way to accomplish this is
>         to let the packets land where they may (the queue where the
>         NIC places the packet) and distribute them (to other queues)
>         by having some of the CPU processing devoted to this task.
>          Can you verify this?
>
>         Regards,
>         - Michael.
>
>
>         On Mon, Jan 6, 2014 at 10:21 AM, Ivan Boule
>         <ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org <mailto:ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>
>         <mailto:ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org <mailto:ivan.boule-pdR9zngts4EAvxtiuMwx3w@public.gmane.org>>>
>         wrote:
>
>             On 12/31/2013 08:45 PM, Michael Quicquaro wrote:
>
>                 Has anyone used the "port config all reta (hash,queue)"
>                 command of testpmd
>                 with any success?
>
>                 I haven't found much documentation on it.
>
>                 Can someone provide an example on why and how it was used.
>
>                 Regards and Happy New Year,
>                 Michael Quicquaro
>
>             Hi Michael,
>
>             "RETA" stands for Redirection Table.
>             It is a per-port configurable table of 128 entries that is
>         used by the
>             RSS filtering feature of Intel 1GbE and 10GbE controllers to
>             select the
>             RX queue into which to store a received IP packet.
>             When receiving an IPv4/IPv6 packet, the controller
>         computes a 32-bit
>             hash on:
>
>               * the source address and the destination address of the
>         IP header of
>                 the packet,
>               * the source port and the destination port of the UDP/TCP
>             header, if any.
>
>             Then, the controller takes the 7 lower bits of the RSS
>         hash as an
>             index
>             into the RETA table to get the RX queue number where to
>         store the
>             packet.
>
>             The API of the DPDK includes a function that is exported
>         by Poll Mode
>             Drivers to configure RETA entries of a given port.
>
>             For test purposes, the testpmd application includes the
>         following
>             command
>
>                 "port config X rss reta (hash,queue)[,(hash,queue)]"
>
>             to configure RETA entries of a port X, with each couple
>         (hash,queue)
>             contains the index of a RETA entry (between 0 and 127
>         included)
>             and the
>             RX queue number (between 0 and 15) to be stored into that
>         RETA entry.
>
>             Best regards
>             Ivan
>
>             --     Ivan Boule
>             6WIND Development Engineer
>
>
>
>
>     -- 
>     Ivan Boule
>     6WIND Development Engineer
>


-- 
Ivan Boule
6WIND Development Engineer