Re: [bmwg] Meeting Minutes Review: IPsec Terminology

[Tim Van Herck <[e-mail]>]

	Brian,

>>>1. Imix is a strict defined packetmix
>>>   - poses problems with newly introduced links and also IPv6, it
>>>     will work with old data
>>>   - has the advantage of providing an apples-to-apples comparison 
> 
> This does provide apples-to-apples comparison of device performance but as
> has already been suggested, one of the primary drivers for an IMIX number is
> for marketing/sales ease of use.  And as such the number is not just used to
> compare one vendor to another but also to provide some notion of customer
> experience for a given service.  So, the strictly defined mix does start to
> become apples-to-orange from the perspective of trying to understand how a
> given device will perform for a given application.  Of particular note in
> this discussion is for VPN applications, for at least in my experience the
> Ipsec devices that we benchmark are used to build VPNs and to forwad data
> for enterprise  applications securelty across the Internet and not merely to
> send generalized Internet data through an Ipsec tunnel.  

It is indeed correct that traffic over an IPSEC tunnel will not be the 
same as generalized Internet traffic since more emphasis will be on 
enterprise client-server application generated data as described in EMIX.

>>>2. Imix is -a- mix, user reports what it is
>>>   - hard to compare results
>>>   - probably will still use the old style IMIX distributions 
> 
> If IMIX must be defined at all this would probably get my vote as it
> provides the freedom to create custome mixes that suite the application of
> the device.  Now my perspective is that of the service provider who is
> characterizing service expectations and it is not the perspective of an
> equipment manufacturer, who could conceivably create a mix that is favors
> the nature of their device rather than any particular application (we
> already see vendors using simple IMIX streams in the VPN space that omit the
> 1500 byte packet in favor of a 1400 byte packet because it exposes the
> inefficiency with which they reassemble fragments). So this is an admitted
> down side but I don't see a more positive approach other than perhaps to
> standardize many different mixes suitable for a variety of applications, but
> I don't see this as being feasible.

The more we discuss about this the more I also tend to go with this 
solution where we only loosely define IMIX as -a- mix and suggest a few 
types but make sure that during reporting the mix is described since 
this will highly influence the results.

Next we will also stress that this is an optional test and that this is 
not so much of a engineering style test but rather a marketing gimmick 
that it used and that it is not a prefered traffic stream for testing 
(as Scott suggested). But I still feel we need to get the industry in 
line about IMIX since the term is used often without any reference to 
what is behind it and as you just mentioned IMIX can take all forms and 
shapes at present.

>>>3. Imix is a packetsweep (after renaming the mix)
>>>   - good chance of comparable results
>>>   - can be adjusted to the needs of the network
>>>   - exercises the full datapath
>>>   - more futureproof
> 
> One of the downsides to this is that you will essentially create a strictly
> defined packet mix, with all of it's downsides.  For instance, if as
> suggested you sweep from min to max and do so in fixed n-byte increments,
> you may end up yielding results that unfairly favor one vendor over another
> simply based upon the nature of their internal architecture, which may for
> instance use n-byte cells to move data chuncks through the switching fabric.
> Modify the sweep to n+1 increments and you get vastly different results.  So
> how do you choose n?  There are potential solutions, like randomizing the
> packet size, etc., but ....

And then the discussion over which packetsizes to include start all over 
again. So this is not an option ! A sweep with 1 byte increments would 
solve the architecture dependency but it will have to be linked to the 
MTU and that again changes the average packetsize of the mix which then 
creates incomparable results all over.

>>>Let the voting begin !
> 
> ... I still vote for number 2, simply define the general concept of a mixed
> packet stream (in a general benchmarking terminology document so that the
> definition carries across all BMWG efforts) and require that the stream be
> clearly defined to include not only packet size distribution but the
> interleaving pattern and any other parameters that may further characterize
> the nature of the stream.

We'll do that ! Unless anybody has another opinion about it ...

	Thanks,
	Tim-


_______________________________________________
bmwg mailing list
[e-mail]
https://www1.ietf.org/mailman/listinfo/bmwg