Hi Martin, Thank you so much for the review! The authors (Luis, Sabine, ...) just had a review meeting this morning and your review is very timely.
Please see below. On Fri, Apr 17, 2020 at 3:45 PM Martin Duke <martin.h.d...@gmail.com> wrote: > This is not a full review. I'm late to this and I apologize if I'm > covering ground which has been fully resolved before. With all hats off... > > It has come up in the working group before that some of the cost metrics > can vary quite a lot over time, much more quickly than the scale at which > ALTO would update them. > > Have we considered changing the metrics to have more long-lived properties > while still being useful? For example > 1. we could split the various delay metrics into min and max versions. Min > and Max delay should be relatively constant over a given topology. Max > delay maps quite nicely to something like an SLA. > Indeed. We discussed more stable statistics (e.g., min can be stable if it reflects propagation delay and no major reroute, 95-percentile can be more stable, ...). > 2. packet loss -- can we focus this on link layer losses rather than queue > drops? Most simply, this could just be a boolean that says "this path > frequently has losses not related to congestion" but could also be expanded > into a link packet error rate, if valuable. Congestion based losses are of > course dependent on the application. > This is an interesting point! I am not an expert on many link layer technologies, but I assume that the loss (more like error) rate is more stable in many settings. Early this week during the NSF Huge Data workshop, one discussion item was that we are approaching a new state where many more files can be corrupted but not detected by TCP checksum for large-scale data transfers. It is not clear where the corruption happens (link link, a buggy switch, ...), but I assume that it is more stable. It is a good suggestion. Back up at a slightly higher level, one key issue that we want to resolve before we are finally happy to wrap up this document is whether we can reduce the number of metrics, instead of having a plethora of metrics: set of metrics (delay, loss, bandwidth, ...) X set of statistics (min, max, percentile, avg) X set of measurement settings (ping each x sec) >From your comments, one direction I see is to define a small subset of the above cross-product space, as a first step. One we define a larger set but make sure to include the more stable/more-likely-quickly-usable settings? Thanks again for the timely feedback! Richard > Thanks > Martin > >
_______________________________________________ alto mailing list alto@ietf.org https://www.ietf.org/mailman/listinfo/alto