Re: [dpdk-dev] [RFC] Add RIB library

[Thomas Monjalon <thomas@monjalon.net>]

Bruce, Vladimir,
There was no progress since August.
Is there a plan to benefit from Vladimir's work?

15/08/2017 13:01, Vladimir Medvedkin:
> Moreover rte_rib_v4_node could contain app specific extension (.ext field).
> For example you can implement PIC (prefix independent convergence) by
> linking rte_rib_v4_node with similar next hop together and precalculate
> feasible next hop for each. Something like:
> struct rte_rib_pic_nh {
>     struct *rte_rib_v4_node;
>     uint64_t nh;
>     uint64_t feasible_nh;
> }
> and keep that linked list's head in next hop structure.
> When next hop fails you just jump from rte_rib_v4_node rte_rib_v4_node and
> change next hop very fast.
> 
> 2017-08-15 13:49 GMT+03:00 Vladimir Medvedkin <medvedkinv@gmail.com>:
> 
> > The advantage is in increasing control plane operations speed. I tested
> > with my fullview + internal routes. It had 650030 prefixes(shuffled) with
> > 1000 specific (longer /24) prefixes within 73 /24 networks. Every prefix
> > had unique next hop. In this test the speed of inserting new routes was
> > increased 70 times against current LPM. This was achieved due to
> > 1. keeping routes in a trie structure instead of array (no need to get
> > free room for rule)
> > 2. avoid unnecessary reads in tbl24 (checking for .depth). Thanks to
> > rte_rib_v4_get_next_child() (that is reverse order traversal without
> > recursion) you can get all more specific prefixes inside your target prefix
> > (you want to add/del), so you can get all ranges between that more
> > specifics and write next hop unconditionally to tbl24 and tbl8.
> >
> > 2017-08-15 11:23 GMT+03:00 Bruce Richardson <bruce.richardson@intel.com>:
> >
> >> On Tue, Aug 15, 2017 at 01:28:26AM +0300, Vladimir Medvedkin wrote:
> >> > 2017-08-14 13:51 GMT+03:00 Bruce Richardson <bruce.richardson@intel.com
> >> >:
> >> >
> >> > > On Tue, Jul 11, 2017 at 07:33:04PM +0000, Medvedkin Vladimir wrote:
> >> > > > Hi,
> >> > > >
> >> > > > I want to introduce new library for ip routing lookup that have some
> >> > > advantages
> >> > > > over current LPM library. In short:
> >> > > >      - Increases the speed of control plane operations against lpm
> >> such
> >> > > as
> >> > > >        adding/deleting routes
> >> > > >      - Adds abstraction from dataplane algorythms, so it is
> >> possible to
> >> > > add
> >> > > >        different ip route lookup algorythms such as
> >> > > DXR/poptrie/lpc-trie/etc
> >> > > >        in addition to current dir24_8
> >> > > >      - It is possible to keep user defined application specific
> >> > > additional
> >> > > >        information in struct rte_rib_v4_node which represents route
> >> > > entry.
> >> > > >        It can be next hop/set of next hops (i.e. active and
> >> feasible),
> >> > > >        pointers to link rte_rib_v4_node based on some criteria (i.e.
> >> > > next_hop),
> >> > > >        plenty of additional control plane information.
> >> > > >      - For dir24_8 implementation it is possible to remove
> >> > > rte_lpm_tbl_entry.depth
> >> > > >        field that helps to save 6 bits.
> >> > > >      - Also new dir24_8 implementation supports different next_hop
> >> sizes
> >> > > >        (1/2/4/8 bytes per next hop)
> >> > > >
> >> > > > It would be nice to hear your opinion. The draft is below.
> >> > > >
> >> > > > Medvedkin Vladimir (1):
> >> > > >   lib/rib: Add Routing Information Base library
> >> > > >
> >> > >
> >> > > On reading this patch and then having discussion with you offline, it
> >> > > appears there are two major new elements in this patchset:
> >> > >
> >> > > 1. a re-implementation of LPM, with the major advantage of having a
> >> > > flexible data-size
> >> > > 2. a separate control plane structure that is designed to fit on top
> >> off
> >> > > possibly multiple lookup structures for the data plane
> >> > >
> >> > > Is this correct?
> >> > >
> >> > Correct
> >> >
> >> > >
> >> > > For the first part, I don't think we should carry about two separate
> >> LPM
> >> > > implementations, but rather look to take the improvements in your
> >> > > version back into the existing lib. [Or else replace the existing one,
> >> > > but I prefer pulling the new stuff into it, so as to keep backward
> >> > > compatibility]
> >> > >
> >> >
> >> > > For the second part, perhaps you could expand a bit more on the
> >> thought
> >> > > here, and explain what all different data plane implementations would
> >> > > fit under it. Would, for instance a hash-lookup work? In that case,
> >> what
> >> > > would the data plane APIs be, and the control plane ones.
> >> > >
> >> >
> >> >  I'm not sure for _all_ data plane implementations, but from my point of
> >> > view compressed prefix trie (rte_rib structure) could be useful at least
> >> > for dir24_8, dxr, bitmap handling. Concerning to hash lookup, it
> >> depends on
> >> > algorithm (array of hash tables indexed by mask length, unrolling
> >> prefix to
> >> > number of /32).
> >> > Perhaps it is better to waive the abstraction and make LPM as primary
> >> > struct that keeps rte_rib inside (instead of rules_tbl[ ]).
> >> > In that case rte_rib becomes side structure and it's API only for
> >> working
> >> > with a trie. LPM's API remains the same (except next_hop size and LPM
> >> > creation).
> >> >
> >> >
> >> What is the advantage of using the rte_rib for control plane access over
> >> the existing rules table structure. Are not the basic operations needed
> >> for adding/removing/looking-up rules supported by both?
> >>
> >> /Bruce
> >>
> >
> >
> >
> > --
> > Regards,
> > Vladimir
> >
> 
> 
> 
>