Hi David,
We will add more text to clarify rationale for L2_OVERHEAD.
Do you have a preference or suggest a way to specify L2 rates? We will work on defining 2 TSpecs to specify two rates (min and max).
As far as VPN terminating between Consumer and Producer, I am not aware of use-cases, at least in the context of the deployment scenario considered for the draft.
Regards,
Shitanshu
From: Black, David <David.Black@xxxxxxxx>
Sent: Monday, February 27, 2017 10:42 AM
To: Shitanshu Shah; tsv-art@xxxxxxxx
Cc: idr@xxxxxxxx; draft-ietf-idr-sla-exchange.all@xxxxxxxx; ietf@xxxxxxxx; Black, David
Subject: RE: Review of draft-ietf-idr-sla-exchange-10
Ok, please make that rationale for the L2_OVERHEAD parameter clear in the draft:
> Though for the cases where physical links are not to be over-run, and/or, sla defined are based on l2 rates, the Consumer would
> need to know l2 overhead from the Producer. Otherwise 10 byes of l2 overhead difference on 64 bytes packets can be significant
> compare to on 1400 bytes packets, and thus the Producer does not have a way to convert rates to ip based (this can cause
> functional issue), since this has to be per packet consideration.
It looks like the result is that if the Producer’s L2 overhead is larger than the Consumer’s, and the token buckets are specified in IP octets, the consumer
has to calculate a per-packet charge of the L2 overhead difference against the IP token bucket (and in the opposite case, there may be a per-packet credit). Are there cases where the overhead is not all L2, e.g., SLA/TCA Consumer is using a VPN that terminates
between Consumer and Producer?
From: Shitanshu Shah [mailto:shitanshu_shah@xxxxxxxxxxx]
Sent: Friday, February 24, 2017 1:05 PM
To: Black, David; tsv-art@xxxxxxxx
Cc: idr@xxxxxxxx; draft-ietf-idr-sla-exchange.all@xxxxxxxx; ietf@xxxxxxxx
Subject: Re: Review of draft-ietf-idr-sla-exchange-10
Hi David,
One response inline ##svshah2
David> Inline ...
Hi David,
Thank you for taking time for the review..
Please find our response inline ##svshah and [Med]
Regards,
Shitanshu
Reviewer: David Black
Review result: Not Ready
I've reviewed this document as part of the transport area
directorate'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.
Document: draft-ietf-idr-sla-10
Reviewer: David Black
Review Date: February 21, 2017
Review result: Not Ready
This is an early TSV-ART review of a working group draft, requested by
the IDR Working Group.
This draft defines an extension to BGP to allow exchange of traffic
handling parameters (e.g., configured rates, burst sizes, drop
thresholds). While this is a useful area of technology to standardize
across network operators, this draft has significant problems, and
parts of it could use some serious rethought.
This reviewer has discussed small portions of this draft with some of
the authors in the past, but this is his first comprehensive reading
and review of the draft.
Major Issues:
[1] The draft is misnamed. This is not an SLA (Service Level
Agreement) draft - it's a TCA (Traffic Conditioning Agreement) draft -
see the definition of TCA in RFC 2475. This draft should start from
that definition and generalize the applicability of TCA beyond
Diffserv. Large areas of SLA content are not covered by this draft -
for more details, see the Wikipedia article on SLA:
https://en.wikipedia.org/wiki/Service-level_agreement .
##svshah, sure, we can change the term to TCA (Traffic Conditioning Agreement), which fairly represents intention in the draft, if there is no otherwise comment from the working group.
David> That would be good, thanks.
[2] Section 3.3.2.1's reuse of the TSpec construct from RFC 2115 to
specify a token bucket is a good idea, but it's not a good idea to
respecify that construct in terms of L2 (link-layer, e.g., Ethernet)
octets, as the RFC 2115 TSpec is specified in terms of IP octets.
This change to specification in terms of L2 octets results in needing
the L2_OVERHEAD TLV to cope with the possible differences in L2 (link)
framing overhead at sender and receiver of this information. The
TSpec should be respecified at the IP layer, with L2 framing overhead
left to the Producer and Consumer to factor into their calculations
based on each's direct knowledge of L2 functionality and configuration
in the local AS. That ought to enable elimination of the L2_OVERHEAD
TLV, thereby reducing complexity.
##svshah, the purpose for advertising L2 overhead, from the Producer to the Consumer, is to make sure traffic is well conditioned, at the egress of
the Consumer, to avoid possible indiscriminate drops at the ingress of the Producer. In another words, the Producer is telling the Consumer to ignore its own link level overhead, instead use Producer's provided link level overhead while running frames through
QoS functions (this is relevant in use-cases where l2 overhead of the egress link on Consumer and of ingress link on the Producer are of different size, e.g., in the vpn/tunnel connection between two peer nodes which physically may be multiple hops away).
I understand your comments regarding re-using TSpec without any modification. Do you agree with the use-case of l2 differences? If yes, any suggestion
how to accommodate that?
David> Specifying this in terms of IP octets implies that L2 overhead is to be ignored on both links. That’s simpler, and a few sentences of discussion
about possible L2 framing differences should cover what needs to be done (e.g., if traffic rates are configured in terms of L2 octets, then explain how each entity converts between IP traffic and L2 traffic - the draft already effectively contains that explanation).
##svshah2, just to make sure that some of the real use are not left out..
For the traffic conditioning where sla established is for the ip based rates only, it suffices for the Producer to send rates only IP based ignoring
any specific of l2 overheads.
Though for the cases where physical links are not to be over-run, and/or, sla defined are based on l2 rates, the Consumer would need to know l2 overhead
from the Producer. Otherwise 10 byes of l2 overhead difference on 64 bytes packets can be significant compare to on 1400 bytes packets, and thus the Producer does not have a way to convert rates to ip based (this can cause functional issue), since this has
to be per packet consideration.
Regards,
Shitanshu
[3] A token bucket should have two rate-related marking parameters
based on its token fill rate, i.e., min-rate, not the four
rate-related marking parameters in this draft. The max-rate
parameters in sections 3.3.2.5-6 ought to be specified against a
second token bucket. In addition, the handling precedence algorithm
in section 3.3.2.7 is an overly complex way to specify the
relationship of two token buckets. All of this is even more
important, because max-rate, as defined in RFC 2115, is only
applicable to bursting - that max-rate for bursting often turns out to
be an interface line rate, which is not generally useful for the
traffic provisioning purposes of this draft.
##svshah, what you describe below conceptually very much makes sense and that is what we attempt to achieve. What is unclear though how to capture
that using TSpec definition specified in RFC2215. Since that TSpec definition has both minimum-rate and maximum-rate, but no in/out profile marking parameters.
Is your proposal below to define two TSpecs using the same definition from RFC2215? Wouldn't in that case we have min/max specified twice? min/max
in each TSpec? and thus confusing to represent one min and one max through two TSpecs?
David> See RFC 2698 for the desired behavior and the parameters needed to specify it. The current idr-sla draft is not able to represent the traffic
conditioning mechanisms specified in both RFC 2698 and RFC 2697, as each mechanism requires two token buckets (e.g., there are two burst sizes, but the idr-sla draft TSpec only contains one). This deficiency needs to be dealt with. One possibility for simplification
is to drop the max-rate from the RFC 2115 TSpec and assume that the max-rate for bursting is the effective line rate.
The following should be done instead:
- Define TSpecs for two token buckets, a primary/committed token
bucket
and a secondary/peak token bucket that MUST be nested, i.e.,
traffic
that is in-profile for the secondary/peak token bucket is always
in-profile for the primary/committed token bucket. Some of the
details
of how to specify this are subtle, see RFC 2698 for a worked
example.
Use of a secondary/peak token bucket requires use of the primary/
committed token bucket, but a primary/committed token bucket can be
used
without a secondary/peak token bucket.
- For a single token bucket, define two handling TLVs, Committed (in
profile) and
Excess (out of profile).
- For two token buckets, define three handling TLVs, Committed (in
profile for both
token buckets), Peak (out of profile for primary/committed token
bucket, but in
profile for secondary/peak token bucket) and Excess (out of profile
for both
token buckets).
NB: Could use Green/Yellow/Red terms instead of Committed/Peak/Excess
terms.
[4] The drop threshold TLV in section 3.3.2.8 is not specified
sufficiently to be implemented interoperably. For example, I don't
understand what an implementation is supposed to do when it receives 3
drop thresholds.
##svshah, It is around the semantics of a single queue with a different threshold for a set of code-points, where packet for a specific code-point
is to be tail dropped if overall queue-depth hits code-point specific threshold at the arrival of that packet.
We will add appropriate clarification for this semantics.
David> Will look at revision
[5] The relative priority TLV in section 3.3.2.9 has the same
insufficient specification problem as the drop threshold TLV,
compounded by a functional incompleteness problem - if the recipient
is using a weighted packet transmission scheduler (e.g., WRR),
priorities cannot be used to configure that scheduler. Hence, some
specification of weights and scheduling algorithms that use weights
needs to be added.
##svshah, Sure. Is following clarification okay? (some of the wordings taken from RFC2598)
A higher priority class of traffic to be served without pre-empted by lower priority class of traffic for more than a packet time at the configured
rate.
In the system that implements WRR, the use of relative priority may get restricted where a single queue may be used for a higher priority traffic class
where that queue is configured for the full share of the output bandwidth.
David> Last sentence basically says that WRR weights are out-of-scope. That seems wrong. Will look at revision.
[6] I have no idea what the sub-traffic classes TLV in section
3.3.2.10 is supposed to do, as that TLV is specified based on "Traffic
Class TLVs" which is an undefined term in this draft (e.g., that term
is not used outside of section 3.3.2.10.
##svshah, They are to facilitate hierarchy. A specific Traffic Class, can further be divided in a multiple subset of Traffic Classes, with their own
Traffic Class Elements and Traffic Class Services.
Will correct the wording to remove your this specific concern.
David> Will look at revision
[7] This draft's QoS contents need to be functionally aligned with
work-in-progress on YANG QoS models, in order to provide some
assurance that that this draft is implementable for actual network
switch/router data paths. The current acknowledgement of the
existence of YANG, NETCONF and RESTConf at the end of Section 1 does
not suffice.
##svshah, While eventually "Traffic Conditioning Agreement" should be translated to the actual forwarding qos policy on any vendor specific device,
TCA exchange largely carries concepts/semantics that is either standard based or well understood.
While we are not aware of any working group document of QoS Yang Model, we think that concepts of Traffic Conditioning should be easily adaptable
to any QoS Yang Models since they also have to be defined to support those concepts.
David> Still want to see cross-check with implementations here.
[8] There are significant complexity and correctness problems caused
by the option to not specify the Source AS - e.g., Section 3.2 defines
an SLA ID as an "identifier which is unique in the scope of Source AS"
which is meaningless if there is no Source AS. It would be simpler to
always specify Source AS, even in the point-to-point case.
##svshah, okay. Ron Bonica also suggested same in his review earlier. We had chosen a path of optional Source AS to simplify implementations in cases. Given specification of Source AS always is more clearer in multiple feed-back we have gotten so far, we can
modify the specification to incorporate that over the simplicity of implementation for certain cases.
[9] In section 3.2, the "intended for the peer receiver of the BGP
UPDATE message" text in the specification of bit 0 of the SLA Subtype
flags is unclear. I suspect that this is intended to differentiate
the two usages described in sections 4.1.1 (Point-to-Point) and 4.1.2
(Multiple Hops), in which case the parenthesized terms (or similar
terminology) should be used with cross-references to those two
sections.
##svshah, sure will make necessary changes, also in the context of earlier comment.
[10] There needs to be a coherent discussion in one place about how
SLA advertisement, update and withdrawal work. A single ADVERTISE
method may suffice on the wire, but the details on how initial
advertisement, subsequent advertisement (update) and withdrawal work
need to be specified in one place. The third paragraph of Section 4
is a start on this material, but it's too terse;
[Med] That text is terse because we are reusing the BGP machinery for advertising/withdrawing; we only augment it with triggers that are linked to the SLA information.
A new SLA will lead to an update message with ADVERTISE, an update of an existing SLA will trigger an update message. We do think this information is already present in the document.
David>This is hard to understand it, as there is no statement that the BGP machinery is being used, and the material on SLA usage is spread in different
places. I suggest starting section 3 with a discussion of advertisement/update/withdrawal, and how all three are realized via a single ADVERTISE method via BGP UPDATE - this was difficult to figure out in reading the current draft.
it should be expanded
into its own subsection, and moved earlier to come before the
ADVERTISE method in Section 3.2. This text from Section 3.2 should be
moved into that new subsection and likewise expanded:
[Med] Another option is to move it Section 4. From our standpoint, we do think it is straightforward to describe the behavior once the attributes format are defined.
David> Ok, that’s an editorial choice - I prefer to explain how something works before defining its on-the-wire data format.
If an advertised SLA ID is different from earlier advertised
one,
for the same prefix and from the same Source AS, indicates
Source
AS is advertising new SLA Content to replace the previous one
advertised with the same SLA ID.
In addition, I wonder whether functionality should be added to allow
withdrawal of an advertisement by specifying its SLA ID, although that
was not part of the original design.
[Med] The reason we didn’t adopted that design is that we don’t want to make assumptions on which data the remote
peer will be used for enforcing local actions and also because of this text:
The SLA ID applies to aggregate traffic to prefixes for a given
AFI/SAFI that share the same Source AS and SLA ID.
David> That’s fine, I wanted to ensure that this had been considered. The discussion of BGP mechanism reuse will probably help
here.
[11] Notions of context for interpretation of all the IPFIX parameters
in 3.3.1 need to be added, e.g.:
- The first three parameters (DSCP, MPLS EXP field in top label,
802.1q priority) can
and do vary on a link-by-link or LSP-by-LSP basis along a traffic's
network path.
- The IP address parameters are rather likely to be VPN-specific when
there's more
than one BGP/MPLS VPN that spans or transits the ASs involved.
- The transport port parameters need specification of which transport
header and where it
is located (e.g., for TCP traffic carried by in VXLAN, is this the
inner TCP header
or the outer UDP header in VXLAN).
There are probably simple approaches to specifying context in all
cases, but that context does need to be specified ... in all cases.
[Med] We dind’t include a context field because we thought that the definition of the IPFIX attributes is sufficient by itself, but if you do think it is helpful to have such information,
we can update the table.
David> Really??? Please reread the first comment above:
- The first three parameters (DSCP, MPLS EXP field in top label, 802.1q priority) can
and do vary on a link-by-link or LSP-by-LSP basis along a traffic's network path.
David> Taking just DSCP, it can change on a per-link basis, so something has to specify which link (between the SLA Producer and SLA Consumer?)
is involved. It probably suffices to say that it’s the link that crosses the relevant AS boundary, but that does have to be stated, as it’s nowhere to be found in the far more general IPFIX registry.
[12] The security considerations (section 10) are severely incomplete
and insufficient:
David> I stand by this statement and recommend a complete rewrite of the security considerations.
- Discussion of possible abuse of this BGP option for
denial-of-service and theft-of-service,
needs to be added, including possible countermeasures and
mitigations.
[Med] This attack vector is not specific to this attribute; this is valid for BGP in general.
David> There are additional attacks enabled by this attribute.
- This sentence at the end of the second paragraph in Section 10 is
content-free:
It is NOT RECOMMENDED to enable this attribute at the
scale of the Internet unless if means to prevent leaking
sensitive
information are enforced.
What exactly is an implementer or admin supposed to do?
[Med] An example of what an admin needs to check if this information can be sent to a given peer or not. Some of the content of the attribute contains some
sensitive information such as contract Id, classes, etc.
How does
one
figure out whether an implementation or deployment is "at the
scale
of the Internet" ??
[Med] The idea here is to avoid leaking such data in to the global routing table. Only entitled peers need receive such information with controlled scope.
David> I understand the idea - the problem is that I don’t see specification of what has to be implemented in either the text
in the draft or the above explanations.
- The next to last paragraph in section 10 is almost content-free, as
it leaves
decisions on implementation and deployment of key security
functionality as
"an exercise for the reader" - that's not acceptable.
[Med] I guess you are referring to the following text. Validation checks are really deployment-specific. The text calls out the issue and
recommends an action; how to translate this action into detailed actions is realty local to a domain
The attribute may be advertised by a misbehaving node to communicate
SLA parameters that are not aligned with the SLA agreements. Though
the enforcement of SLA parameters is outside the scope of this
document, it is RECOMMENDED that the SLA Consumer to enforce a set of
validation checks before translating the SLA parameters conveyed in
the QoS attributes into provisioning actions. Such validations MAY
rely on SLA parameters like the origin AS or SLA ID, like generating
SLA ID using pseudo-random schemes [RFC4086].
David> I stand by the “exercise to the reader” criticism - one way to be honest about this would be to change the paragraph to:
The attribute may be advertised by a misbehaving node to communicate
SLA parameters that are not aligned with the SLA agreements. The
enforcement of SLA parameters is outside the scope of this document.
David> That does not change any normative content of the original paragraph. My view is that the “outside the scope of this document” statement is wrong because all of these parameters are being defined in this draft.
- Last, but not least, the final paragraph in Section 10 is a joke
that will
be lost on a security directorate reviewer - to understand why, see
Section 2 of RFC 6919, and take note of the publication date of RFC
6919.
David> For anyone who didn’t catch the reference, RFC 6919 is an April 1 RFC ... and the “SHOULD BE considered” language used in the last security considerations paragraph of this draft is accurately characterized as vague in Section 2 of RFC 6919:
The phrase "SHOULD CONSIDER" indicates that the authors of the
specification think that implementations should do something, but
they're not sure quite what.
------------------------
Having noted a dozen major issues, I'll end the review here for now,
as I believe a serious revision of the draft is called for, which
would be a better starting point to review for minor issues and
editorial items.
--------------------------------------------------------
David L. Black, Distinguished Engineer
Dell EMC, 176 South St., Hopkinton, MA 01748
+1 (508) 293-7953 Cell: +1 (978) 394-7754
David.Black@xxxxxxxx <=== NEW ===
--------------------------------------------------------
|