Re: [Sidrops] WGLC = draft-ietf-sidrops-aspa-verification - ENDS 03/22/2023 (Mar 22 2023)

[Claudio Jeker <[e-mail]>]

On Tue, Mar 28, 2023 at 05:54:26PM +0900, Ben Maddison wrote:
> Hi Claudio,
> 
> On 03/23, Claudio Jeker wrote:
> [..]
> > I have a major issue with duplicating the CAS lookup table. The
> > implication for BGP implementations is twice the memory consumption and
> > more then double the computational cost and that in the most latency
> > sensitive portion of the code. A lot of time is spent in getting the most
> > performance out of BGP and so this should be a major consideration when it
> > comes to ASPA validation.
> 
> Would you elaborate on this?
> 
> I get why the memory usage increases with this approach.
> 
> It's not obvious to me why the computational cost of the verification
> process is higher with a per-AFI table.
> Intuitively this seems like it should be a *cheaper* lookup, since you
> don't have to filter for AFI at lookup-time.

The problem is that the tables are close copies of each other.
So you end up needing twice as much cache to handle the same amount of
information. Since cache is king in todays world of high speed CPUs the
fact that twice as much is needed is a problem. Data may no longer fit
and so expensive roundtrips to main memory are needed.
Modern computers do not behave like the CS models behind things like the
Big O() notation.
 
In OpenBGPD we additionally calculate and cache all 4 possible results
in one go. There is little overhead in doing the lookup for both AFI since
our hop functions returns both results in a single lookup.
Doing both downstream and upstream checks adds extra lookups but right now
our code has no prior knowledge to skip these if the path is only used by
one group of roles.

-- 
:wq Claudio