Re: [alto] Tsvart early review of draft-ietf-alto-performance-metrics-03

["Brian Trammell (IETF)" <[e-mail]>]

On further review, I have an additional question about the metrics specified in the document:

Max reservable (sec 8.1) and residue (sec 8.2) bandwidth seem to be fairly straightforward to calculate in a stable manner, as they are essentially matters of router configuration, but I'm a little concerned about the temporal aspects of available and utilized bandwidth. On any link containing congestion controlled traffic not otherwise limited (e.g. by a tunnel reservation), even a small number of active bulk transfer flows will cause the utilized bandwidth to max out on short timescales: you get an almost uselessly noisy metric.

I presume, since this metric is taken straight from an IS-IS standards track RFC, that this doesn't happen in practice, i.e. that the temporal denominator is large enough that it averages these smaller peaks out. It'd be nice if the document named a timescale though.

I'm also a little dubious that instantaneous bandwidth is a useful cost metric for a system like ALTO, because even on sane timescales the delay imparted by the system makes  the information hard to act on in a useful way, unless the link is a large enough aggregate that the bandwidth doesn't change much over time.

Cheers,

Brian

> On 9 Apr 2018, at 18:35, Brian Trammell <[e-mail]> wrote:
> 
> Reviewer: Brian Trammell
> Review result: On the Right Track
> 
> I've performed a (late, apologies) early TSV-ART review of
> draft-ietf-alto-performance-metrics-03.
> 
> The set of metrics chosen by the document seem broadly useful and sane, and the
> integration into ALTO makes sense. However, there are a few issues with the
> details.
> 
> Periodic One Way Delay, RTT, and PDV are defined in terms of section 8, section
> 4, and section 5, respectively, of draft-ietf-ippm-initial-registry, which
> specify active measurement test methodologies at layer 4 for one-way and
> round-trip delay using UDP packets. This does not seem it can be measured
> directly using the routing  technologies the authors have identified as their
> source of information. Is the intention that dedicated active measurement
> hardware be used to measure delay using UDP packets, or should these metrics
> reference [RFC2679] and [RFC2681] and leave the methodology undefined, instead?
> 
> The examples for these don't make much sense: the units are expressed in
> seconds, but Internet-scale delays are generally millisecond-scale, and the
> examples given contain only integers. Similarly, packet loss rate is given in
> percentile, but there are wide variations in usability between a path with 0%,
> 1%, and 2% packet loss. Is this simply an issue with the examples?
> 
> The hop count metric is underspecified: are these IP-experienced hops at layer
> 3, as can be measured by traceroute?
> 
> Nit: section 2.1 refers to [OSPF-TE], [ISIS-TE], [BGP-LS] and [BGP-PM], but
> these are not listed as such as references in the references section. Please
> use consistent reference labels.
> 
> Thanks, cheers,
> 
> Brian
>