Hi Stig,
I agree with all you wrote and will update the document. However,
there is one slight issue with the minimum time between origination of
each message. When a new source is detected, we would like to
originate a message ASAP so that receivers can start receiving the
multicast without much delay. A 10s delay would be a rather long time
if a source was detected right after the previous message was
originated. I think some delay would be warranted though, in
particular in a case where perhaps a router starts up and a large
number of directly connected sources could be detected within a short
time frame. I think an exponential back-off could make sense here.
E.g., if it is just one new source, maybe trigger a message ASAP. If a
new source is detected right after the previous one, wait a bit
longer, which also allows for aggregation of multiple sources in one
messages if several are detected later. In extreme cases one could
over time keep increasing the delay until the next update.
If sufficient we could maybe have a fixed minimum delay of 1s or not,
but that is probably too short in those extreme cases. Hence maybe an
exponential back-off.
Exponential back-off sounds like a very good idea - I'd suggest adding something starting from RFC 5059's back-off functionality.
I would appreciate some further guidance what you think is reasonable
here, and perhaps whether I can borrow something here from other
protocols/drafts. Part of the experiment here might be to find out
what minimum values, or how rapid back-off, is needed based on the
size of the network, the amount of sources, the types of links etc.
In addition to burst scenarios (e.g., router starts up, lots of new sources detected quickly as a result), I strongly suggest thinking about chaos scenarios where links and/or routers are coming and going so rapidly that the source population is in a constant state of flux. If things are really bad, the best thing to do may be to shut up and hope that the chaos settles out, as not much useful will happen until it does, and send messages about observed changes risks make things worse. Again, exponential back-off makes sense, possibly quite aggressive, e.g., back-off from 10 seconds by a small factor a few times, and if things still look bad, wait at least a minute or two with further back-off from that longer time until things stabilize. This needs more thought on how to adjust the back-off factor, as that off-the-top-of my-head example probably exhibits peculiar behavior in scenarios that just are on the edge of tripping the long delay - some thinking about what stability means and how to get there may help in figuring out the relative merits and applicability of backing off further vs. some kind of dramatic reset, analogous to TCP's congestion window reset on timeout.
As this is intended to be an experimental RFC, I don’t think a completely worked-out solution is expected or required - a good discussion of the problems and explanation of areas that need investigation as part of the experiment ought to suffice, as suggested in last sentence quoted above. I would add some initial exponential back-off functionality as a starting point.
Also note that the general mechanism can be used for many types of
information. It depends on the information how urgent it is to
distribute it. Source discovery is particular is fairly urgent.
And that should be discussed, perhaps in Section 3 somewhere.
Thanks, --David
-----Original Message-----
From: Stig Venaas [mailto:stig@xxxxxxxxxx]
Sent: Tuesday, January 23, 2018 7:44 PM
To: Black, David <david.black@xxxxxxx>
Cc: tsv-art@xxxxxxxx; draft-ietf-pim-source-discovery-bsr.all@xxxxxxxx;
ietf@xxxxxxxx; pim@xxxxxxxx
Subject: Re: Tsvart telechat review of draft-ietf-pim-source-discovery-bsr-08
Hi, thanks for the great comments.
I agree with all you wrote and will update the document. However,
there is one slight issue with the minimum time between origination of
each message. When a new source is detected, we would like to
originate a message ASAP so that receivers can start receiving the
multicast without much delay. A 10s delay would be a rather long time
if a source was detected right after the previous message was
originated. I think some delay would be warranted though, in
particular in a case where perhaps a router starts up and a large
number of directly connected sources could be detected within a short
time frame. I think an exponential back-off could make sense here.
E.g., if it is just one new source, maybe trigger a message ASAP. If a
new source is detected right after the previous one, wait a bit
longer, which also allows for aggregation of multiple sources in one
messages if several are detected later. In extreme cases one could
over time keep increasing the delay until the next update.
If sufficient we could maybe have a fixed minimum delay of 1s or not,
but that is probably too short in those extreme cases. Hence maybe an
exponential back-off.
I would appreciate some further guidance what you think is reasonable
here, and perhaps whether I can borrow something here from other
protocols/drafts. Part of the experiment here might be to find out
what minimum values, or how rapid back-off, is needed based on the
size of the network, the amount of sources, the types of links etc.
Also note that the general mechanism can be used for many types of
information. It depends on the information how urgent it is to
distribute it. Source discovery is particular is fairly urgent.
Stig
On Tue, Jan 23, 2018 at 3:40 PM, David Black <david.black@xxxxxxxx> wrote:
Reviewer: David Black
Review result: Ready with Issues
I've reviewed this document as part of TSV-ART's ongoing effort to review key
IETF documents. These comments were written primarily for the transport area
directors, but are copied to the document's authors for their information and
to allow them to address any issues raised. When done at the time of IETF Last
Call, the authors should consider this review together with any other last-call
comments they receive. Please always CC tsv-art@xxxxxxxx if you reply to or
forward this review.
This draft describes an experimental PFM (PIM Flooding Mechanism) mechanism for
flooding PIM information among multicast routers that is a generalized form of
the RFC 5059 PIM BSR (BootStrap Router) mechanism, and applies this mechanism
to distribution of source group mappings (PFM-SD).
Early implementation experience with PFM-SD on low bandwidth radio links
(described Section 2) suggests that the mechanism is able to work better than
PIM-SM without starving other traffic in the fashion that PIM-DM may. This is
promising and (in this reviewer's opinion) justifies experimentation at larger
scale and in other network environments. In general, this is a well-written
document and the authors should be commended for including the "running code"
implementation experience report in Section 2.
Flooding mechanisms are very useful, but the time periods that govern sending
of flooding messages are crucial to avoid excessive consumption of network
resources. Section 5 of RFC 5059 has a solid discussion of the time periods
that apply to use of flooding by the BSR mechanism. The discussion in this
draft is somewhat weaker, raising a couple of minor issues:
1) For PFM-SD, Section 4.2 provides a reasonable discussion of time periods
that apply, but appears to be missing a minimum time period between sending
messages. Section 5 of RFC 5059 recommends a default of 10 seconds for that
minimum time period by comparison to a default PIM BSR sending interval of 60
seconds. That 10 second minimum default should be added to this draft, as the
same default sending interval of 60 seconds is used.
2) For future use of PFM for other purposes, Section 3.3 provides the following
guidance:
Each TLV definition will need to define when a triggered PFM message needs
to be originated, and also whether to send periodic messages, and how
frequent.
That guidance is correct as far as it goes, but it's not particularly helpful
to future protocol designers. Text should be added to at least point to the
examples in section 4.2 of this draft and/or part of Section 5 of RFC 5059 to
suggest the sorts of values that have proven to be workable, and perhaps also
strongly encourage (SHOULD use) a default minimum time between messages of at
least 10 seconds.
Understanding this draft requires that the reader be familiar with multicast
and PIM, which is reasonable. In addition, an understanding of PIM BSR is also
required, which is perhaps somewhat less reasonable. An example that this
reviewer tripped over is that Section 3 of this draft states that "Like BSR,
messages are forwarded hop by hop." There is no further explanation or
definition of "forwarded hop by hop," making it necessary to consult RFC 5059
to understand that term, e.g., this has nothing to do with IPv6 hop-by-hop
options. A sentence or two of explanation of this hop by hop forwarding
concept ought to be copied and adapted from RFC 5059, and it would be good to
check for other concepts that rely on RFC 5059 for definitions.