Re: Uniform Probabilistic Sampling in PSAMP-PROTO

[Andrew Johnson <andrjohn@cisco.com>]

Juergen Quittek wrote:
> Hi Andrew,
>
> --On 08.12.2005 16:53 Uhr +0000 Andrew Johnson wrote:
>
> > Benoit Claise wrote:
> > [SNIP]
> >
> > > Initially we wanted to model the probability in [PSAMP-PROTO] with a
> > > float, which is allowed by [IPFIX-PROTO].
> > > However, we've got the issue that SMIv2 doesn't support floats.
> >
> >
> > Although the floating point type is not a base type in SMI there are
> > a couple of proposals about how to send them.  This extract is from
> > the comp.protocols.snmp SNMP FAQ:
> >
> >   http://www.faqs.org/faqs/snmp-faq/part2/
> >   ===============================================================
> >   2.40.04
> >   SUBJECT: Floating Point Numbers in SMI?
> >
> >   You cannot add new base types to the SMI.
> >
> >   For now, the easiest way to have floating point numbers
> >   in SNMP MIBs is to use the base type OCTET STRING and
> >   encode the value in ASCII.
> >
> >   This is not the most elegant approach. However, it will work
> >   between your agent and your management application and it will
> >   be compliant to the SNMP SMI and protocol specifications.
> >
> >   David Perkins
> >   ===============================================================
> >
> > Apparently this method is implemented in the NET-SNMP agent and
> > manager.
> >   (http://ops.ietf.org/lists/mreview/mreview.2004/msg00083.html)
> >
> > Whether we use the above suggested method or define a new way of
> > using unsigned32 the application will have to have some way of
> > converting that number before it can be used.
>
>
> This would be Solution 4.
> I still prefer Solution 1 or Solution 2.
>
> > > What to do now?
> > >
> > > _Solution 1: _
> > > We export the probability with a float, and we approximate this value
> > > with the MIB variable.
> > >
> > > _Solution 2: _
> > > We export the probability with an unsigned32, exactly the same content
> > > as the MIB variable psampSampUniProbProbability
> > >
> > > _Solution 3: _
> > > We export the probability with two values N, M.
> > >     This means 2 inter-dependent I.E.s and 2 MIB variables instead of 
> > > one.
> > >     I don't like it too much
> > >
> > > I'm clearly in favor of solution 1. It's not right that we would limit
> > > IPFIX because of the limitations of SMIv2
> > > Feedback?
> > >
> > > Side question: if we go for the float solution, should we have a
> > > float64? This would give us more precision
> > > Note: not yet defined in [IPFIX-PROTO].
> >
> >
> > FYI:
> > The 32-bit float has 24 bits of precision, i.e. roughly +/-0.000006%.
>
>
> It is 23 bits we are using. Precision would be    roughly +/-0.00001%

I think you are correct.  IEEE 764 assumes a top bit of 1 but although
this generates a 24-bit number we only have 23 bits of precission, as
you say.

This precision remains constant as the size of the number varies as the
exponent shifts the mantissa to where the bits have the most accuracy.
This is why I prefer using the float as a solution.


> The Unsigned as Benoit suggests it would have     roughly +/-0.0000001%

The precision scales according to the size of the fraction you are
representing, so for the common small numbers it is quite precise but for
the smaller the fraction the less accurate the method is. For example, for
a 1 in 3 billion sampler the closest we can represent is 1 in 2^32, which
is almost 50% off.


Regards,

Andrew


> > The 64-bit float has 53 bits of precision, i.e. roughly 
> > +/-0.00000000000001%
> > regards
>
>
> Thanks,
>
>    Juergen
>
> > Andrew
> >
> > --
> > to unsubscribe send a message to psamp-request@ops.ietf.org with
> > the word 'unsubscribe' in a single line as the message text body.
> > archive: <http://ops.ietf.org/lists/psamp/>

--
to unsubscribe send a message to psamp-request@ops.ietf.org with
the word 'unsubscribe' in a single line as the message text body.
archive: <http://ops.ietf.org/lists/psamp/>