| TOC |
|
Historically, each Internet application protocol has defined its own set of rules for exchanging control and data information. Although there are a lot of similarities (e.g., dot-stuffing in SMTP and POP, octet counts in HTTP and IMAP, 3-digit reply codes in FTP and SMTP), there isn't a lot of reuse of design.
As a result, designers of new application protocols usually have to make the same set of design decisions — over and over again — to deal with the same set of trade-offs, but usually without the benefit of institutional history.
BEEP is a generic application protocol kernel for connection-oriented, request-response interactions. This particular subset supports a large class of Internet applications, and provides solutions to common design issues for those applications, including: framing, encoding, error reporting, asynchronous exchanges, along with authentication and privacy.
| TOC |
| TOC |
BEEP RFCs 3080, 3081
Other I-Ds:
| TOC |
A lot of application protocols, a lot of similarities, not a lot of reuse
Need an application protocol? Three choices:
If choosing the first option involves little or no work, take it
If the application protocol's requirements are a good fit with the web or email, use it
Otherwise, let's talk about coming up with a decent 90% solution
| TOC |
HTTP has a lot of things to recommend it
The question is how good is the fit with a synchronous, small-request/large-response model
Experience with IPP suggests a subtle shift in existing practice has unintended consequences:
Regardless, the costs of introducing new methods, URL-prefixes, etc., probably neutralize the benefits of reuse
| TOC |
Five operational characteristics:
For example:
| TOC |
Limiting the problem domain:
This doesn't accommodate things like:
That's OK... we can define alternative frameworks later, according to the demands of the standards marketplace
| TOC |
There are (at least) six tasks that application protocols have to perform:
There are other important (application-specific) tasks, e.g., naming and authorization
| TOC |
Three Approaches:
Octet-stuffing tends to be slower due to comparison overhead
Octet-counting, in the absence of segmentation, requires the entire message be available prior to sending
Connection-blasting, for small messages, is high overhead; regardless, endpoint-addresses must be conveyed at the application layer
| TOC |
Quite a few to choose from:
There are arguments for and against each
Sometimes we have to just pick one and move on
| TOC |
Historically, there is consensus on SMTP's "theory" of reply codes:
Some successes at having the same code mean the same thing in different protocols
Typically, paired with an ameri-centric textual diagnostic
There are also weaknesses:
| TOC |
Some success here (the second time around):
Doing asynchrony right requires:
Segmentation allows octet-counters to send segments without the entire message being available
| TOC |
Historically, not a problem, because we didn't do it
Hysterically, if we do it, we do it differently each time (e.g., FTP and HTTP)
The IETF has retrofitted many application protocols to use the SASL framework
SASL solves an NxN problem:
| TOC |
Historically, not a problem, because we didn't do it
Then Netscape gave us SSL (and screwed up by using two ports per application protocol)
The IETF has retrofitted many application protocols to support TLS
(Graphic courtesy of Bob "RL" Morgan, University of Washington)
| TOC |
Mechanism ESMTP FTP HTTP1.1
--------- ----- --- -------
Framing stuffing blasting counting
Encoding 822-style binary MIME
Reporting 3-digit 3-digit 3-digit
Asynchrony pipelining none pipelining/ chunking
Authentication SASL user/pass user/pass
Privacy SASL or TLS none TLS (nee SSL)
| TOC |
Regardless of what the application protocol does, a well-designed protocol is:
| TOC |
Surprisingly few application protocols are network-friendly:
Consider:
| TOC |
Elegance is sometimes inefficient (e.g., octet-stuffing v. octet-counting)
Trade-offs are often necessary:
The SOAP guys are probably right that XML's inefficiencies are offset by printf's simplicity
The key lesson is to adjust the razor based on:
| TOC |
One designer's complexity is another's completeness
The pain should be proportional to the gain:
Two ways of doing the exact same thing suggests poor design
Simplistic doesn't imply simple-minded:
| TOC |
Anticipating the unforeseen is a problem:
Extensibility is no solution to poor design (viz. Telnet)
| TOC |
Postel's Robustness Principle is a two-edged sword:
| TOC |
Mechanism BEEP
--------- ----
Framing counting with a trailer
Encoding MIME, defaulting to application/octet-stream
Reporting 3-digit with localized textual diagnostic
Asynchrony independent channels
Authentication SASL
Privacy SASL or TLS
| TOC |
The header is:
tag SP channel SP msgno SP more SP seqno SP size CRLF
Several "consistency rules" used to validate each frame
The payload is 0 or more octets
The trailer is:
"END" CRLF
So, what's a channel?
| TOC |
A profile defines the messages exchanged on a given channel
At startup, channel 0 is bound to the channel management profile
RPY 0 0 . 0 124 Content-Type: application/beep+xml <greeting> <profile uri="http://xml.resource.org/profiles/TLS" /> <profile uri="http://xml.resource.org/profiles/SEP" /> <profile uri="http://xml.resource.org/profiles/sasl/OTP" /> <profile uri="http://xml.resource.org/profiles/sasl/..." /> </greeting> END <start number="1"> <profile uri="..."> ... </profile> <profile uri="..."> ... </profile> </start> <close number="1" code="200" />
| TOC |
A registration template defines:
Two kinds of profiles:
Protocol =
BEEP + 1 or more profiles
+ authorization policies
+ provisioning rules
Exchanges:
MSG -> RPY
MSG -> ERR
MSG -> zero or more ANS, followed by NUL
| TOC |
All payloads are MIME-encoded, by default:
Content-Type: application/octet-stream Content-Transfer-Encoding: binary
BEEP defines a minimal XML media type, application/beep+xml, i.e.,
| TOC |
Each profile defines messages for both positive and negative replies
For negative replies, the channel management profile uses 3 digit reply codes with a localized diagnostic, e.g.,
<error code="421"
xml:lang="en-US">service not available</error>
A list of preferred languages is passed in the <greeting> when a BEEP session is established
| TOC |
Channels handle asynchrony:
Segmentation is handled by framing
So, what about flow control?
| TOC |
A BEEP session is connection-oriented, and mapped onto a transport service that defines how:
At present, one mapping is defined: onto a single TCP connection
In the future, look for transport mappings onto:
| TOC |
The TCP-based mapping re-introduces the window-based flow control mechanism of TCP, each channel:
When a channel is created, it has a 4K window, after that it's managed:
"SEQ" SP channel SP ackno SP size
BEEP peers expected to use basic fairness algorithms, e.g.,
| TOC |
Each BEEP session has a single authentication/authorization identity:
Start a channel with a SASL profile, e.g.,
<start number="1">
<profile uri="http://xml.resource.org/profiles/sasl/OTP">
<blob>AGJsb2NrbWFzdGVy</blob>
</profile>
</start>
The <blob> is used to carry the binary SASL exchanges using base64 encoding
Once you successfully authenticate:
Note that the SASL mechanism may provide both user authentication and transport security:
| TOC |
Each BEEP session has a single privacy context:
Either start a channel with a SASL profile that has a security layer, or start a channel with the TLS profile, e.g.,
<start number="1">
<profile uri="http://xml.resource.org/profiles/TLS">
<ready />
</profile>
</start>
When the underlying negotiation process begins, all channels are closed
At the end of the negotiation, a new greeting is issued, regardless of the outcome
The usual STARTTLS rules apply, a BEEP peer:
| TOC |
Several applications in varying states of progress:
| TOC |
A client/server relationship with asynchronous callbacks:
C: establish and tune session
C: MSG (tell me about)
S: RPY (here is what i know right now)
...
S: MSG (here is something new)
C: RPY (ok)
| TOC |
A server/client relationship:
C: establish and tune session
S: MSG (i'm ready to act as a sink)
C: ANS (audit entry 1)
C: ANS (audit entry 2)
...
C: ANS (audit entry N)
C: NUL
| TOC |
An interesting requirement:
Realized using a "tunneling" tuning profile:
C: establish session to proxy #1
C: start tunneling profile
C: specify route (via IP or SRV)
P1: establish session to proxy #2
P1: start tunneling profile
P1: specify remaining route (via IP or SRV)
...
Pn: establish session to server
Pn: start pass-through profile
Pn: specify empty route
S: pass back status to Pn
...
If successful:
| TOC |
Application-layer relaying:
Behavior of mesh modified by self-contained options, e.g.,
Starting to look at end-system multicasting
| TOC |
BEEP's core and TCP mapping specifications:
Other IETF activities gearing up
The Working Group: [email protected]
The Website: http://beepcore.org/
Open Source Implementations (in progress):