There are two ways to use EXI to make efficient XMPP communication. The first method describes how to activate EXI-compression using &xep0138; (XEP-0138). The second method describes an alternative binding. This method does not use Stream compression as defined in XEP-0138, rather it allows clients to connect to the server and start using EXI directly from the beginning.

The following sections assume the client connects through the normal XMPP port, and starts communicating with the server using uncompressed XML fragments. When the client connects to the XMPP Server, it will receive a list of features supported by the server:

Support for EXI compression through the normal XMPP port is detected by the existence of the exi compression method in the features stanza. If a port (static or dynamic) is available for a dedicated binary EXI/XMPP binding, this can be detected by the existence of the exi:PORT compression method, where PORT is replaced by the port number used. More information about this alternative method is available in the Alternative Bindings section.

Note: If the client already knows the port number of the dedicated binary EXI/XMPP binding, it can connect there directly, without the need to check the server features using the normal XMPP port.

Following is a list of use cases displaying how the client can configure and activate EXI compression on the current binding.

Becuase EXI engines need to close all open XML elements before decompressing, it cannot start the stream by sending only an open <stream> element, and close the stream by sending a closing </stream> element. Instead separate streamStart and streamEnd elements have to be sent, allowing for similar semantics on the EXI-compressed channel, as described in the following subsections.

For clarity, examples in this section are displayed in XML for readability. But it is understood that the elements are sent using EXI compression and using the options defined during setup.

Alternative binding for EXI/XMPP is suitable for use cases such as factory automation, smart grid appliances, and other embedded use of communications. It works best if clients are constrained and does not update its specification frequently. In addition, the network should allow clients and servers to use not well-known port because this commeunication involves alternative TCP port.

Typical steps of communication is as follows (based on &rfc6120;).

1. (Optional) A client (foo@example.net) try to resolve a server with alternative binding for EXI/XMPP with DNS SRV lookup (ex. _xmpp-bclient._tcp.example.net. IN SRV)

2. (Optional) A DNS server tells a set of DNS RR to notify a server accepts EXI/XMPP binding (ex. SRV 10 10 15222 srv.example.net.) (Optional: the DNS server may tell the version of the default schema supported by the server. Currently there is only one version and has no effect. For further discussion, see draft-doi-exi-messaging-requirement.

3. The client connects to srv.example.net. 15222 with TCP and the server accepts the connection.

4. The client sends out 'EXI Cookie' (e.g. '$EXI') and starts EXI stream with an EXI Header without any option document (implies default encoding parameters). It sends out EXI events corresponds to start tag of <stream:stream>. Following shows XML Equivalent and EXI events

   ]]>

5. The server responds with EXI stream, with EXI cookie, without EXI option, and with appropriate events for stream tag.

   ]]>

6. If client needs TLS or SASL negotiation, it should be done at this step. As specified in Section 4.3.3 of RFC6120, both parties MUST not send events corresponds to </stream:stream> tag. (e.g. exi:streamEnd element)

7. If client needs to use different encoding option or schema than the default encoding options or the default schema, then the client shall start schema negotiation. The streams with alternate options/schemas SHOULD NOT have an EXI Options document to indicate the parameter is negotiated via previous XMPP stream.

   For example, the client want to use MUC option (XEP-0045), the following communication will occur. First, client try to renegotiate XML schema used in the communication.

   ]]>

8. In the response the server accepts schema change.

   ]]>

   Note that the server may deny the negotiation with agreement="false" setupResponse (example omitted).

9. Just after receiving the setupResponse, client re-opens the stream. The new stream should have EXI header and may have EXI option header with updated options.

   ]]>

10. In response, server re-opens the stream with exi:streamStart tag.

    ]]>

11. After an exi:streamStart from the server to the client, they can communicate with EXI stream. The first level element in conventional XMPP is encoded as root element of EXI message. For example, a client may send MUC query with EXI.

    ]]>

    This message has a query element under muc#owner namespace. This is performed efficiently because this series of messages from the last streamStart element has been encoded with the set of schemas and the set inclues schemas for MUC. Otherwise, the encoding will become 'built-in grammar' even if the encoder and the decoder uses non-strict schema-informed grammar. This is not possible if either encoder or decoder does not support built-in grammar or the stream uses strict schema-informed grammar. In such cases, the whole message that contains undefined element or attribute SHOULD be dropped.

12. The client and the server may end the stream with exi:streamEnd tag anytime.

    ]]>

    Exactly same message should be sent from the server to the client to close the opposite way of the stream (example omitted).

13. The client now can start a preconfigured EXI/XMPP stream with explicit EXI option with the encoding option and canoinc schema used in the previous negotiated session (shortcut setup)

    c:761aabc0-a255-4b9b-89a1-4cb859559691
    ]]> ]]>

14. The server may respond with the given configuration ID. This can skip the configuration setup and the communication can continue with preconfigured schemas.

    c:761aabc0-a255-4b9b-89a1-4cb859559691
    ]]> ]]>

    Or server may deny to start the new communication by just sending streamEnd tag with default encoding.

    ]]>

This section describes a common view on programming model between EXI/XMPP servers and clients. This is just an example and an implementation may do it in different ways. However, messages exchanged as a result should be identical.

As shown in this document, XMPP documents represented by EXI is slightly different from documents represented in XML. For example, each stanzas and first level elements under <stream:stream> tags must be encoded as a standalone EXI body (document). If namespace prefixes are declared in the root element of XML process (<stream:stream> tag), the prefix should be recovered on decoded XML on the receiver side (EXI may or may not preserve prefix). Therefore, some preprocessing and postprocessing is required to make current XMPP implementations work properly over EXI/XMPP channel. An example of the configuration is described in the following figure.

In the example, the box XMPP is conventional XMPP process. [EXI] represents EXI processor (encoder and decoder). [Proc] means preprocessor and postprocessor. Following pre-process must be applied to XML streams from XMPP process before giving the stream to EXI processors.

1. <stream:stream> start tag must be converted to standalone <exi:streamStart> element. All attributes defined in the input XML except namespace declarations should be copied as is. Namespace declarations and prefixes must be converted to exi:xmlns element under exi:streamStart element.
2. All first level elements under the root <stream:stream> element must be converted to standalone (root) element of corresponding stanza elements. If there are missing namespace declarations in the element, appropriate prefix declarations should be added to the element before giving the element to the EXI encoder as a document. Contents of the elements should not be modified.
3. </stream:stream> end tag must be converted to standalone <exi:streamEnd> element.

Following post-process must be applied to XML streams from EXI processors before giving the stream to XMPP process.

1. Received <exi:streamStart> element must be converted to <stream:stream> start tag with appropriate attributes and namespace declarations. Prefixies defined in the <exi:streamStart> content must be saved under the current connection context of postprocessor.
2. Received EXI messages must be reproduced as elements under <stream:stream> root element. In addition, the namespace declarations in the element must be replaced to prefixes given in the streamStart tag content if there is a corresponding namespace declaration.
3. <exi:streamEnd> element must be converted to </stream:stream> end tag. The context created under the connection may be released.

There are many EXI options in EXI Format. Followings are brief description for use in XMPP.

The alignment option is to control how the values are encoded. The default option 'bit-packed' fits most of communication use cases. If TLS compression is used at the same time, pre-compression will make the best result.

The compression option is a Boolean to specify additional entropy-based compression is performed over EXI encoding. This option works best for larger documents. Usual use cases expected in XMPP/EXI may not give much additional compactness to messages.

The strict option is a Boolean to let EXI grammars (schema-informed grammars) work more strictly. With a strict schema-informed grammar, only valid data with the schema is allowed in streams. This makes best compactness of a grammar and messages without additional entropy-based compression. With nonstrict schema-informed grammar, derived elements and attributes could be encoded in the built-in grammar. The built-in grammar is a dynamic, self-learning grammar model that gives full flexibility, with larger message size.

The preserve option is a set of Booleans to alter some production of events. Usually, all false (no preservation of comments, processing instructions, DTDs, namespace declarations and its prefixes, and lexical values) makes no problem on XMPP communication. However, uploading a binary schema may require the schema encoded with preserved namespace prefixes. The prefixes are used in type definitions, and without prefix preservation binding between type definition references and actual definitions will be lost.

The selfContained option is a Boolean to enable self contained encoding of elements. Self contained element has no dependency to context (string table state in EXI processors) and can be copied to other context with the same grammar. However, self contained elements cannot use external string tables in EXI processors and may result larger size.

The datatypeRepresentationMap option can be used to modify how types encoded in EXI. If encoders and decoders have special encoding, it can be specified here. In most use cases in XMPP/EXI, this option will not be used.

The blockSize option specifies the block size used for EXI compression. It has no effect if compression is not used.

The valueMaxLength option specifies the maximum length of re-used string in a stream. Larger value makes more strings captured in string tables in EXI processors. This means more memory needs to be allocated to process a stream. Because some use cases, such as Internet of Things, expects constrained nodes in the network, default value specified in ths XEP is very small (64 characters).

The valuePartitionCapacity option specifies how many strings should be kept in a string table in a time. Default value of this is also small in this XEP (64 strings).

To control buffer and string table life time, this XEP adds a new option: sessionWideBuffers. If set to true, all buffers, string tables, etc. will be maintained during the entire session. This may improve performance during time since strings can be omitted in the compressed binary stream, but it might also in some cases degrade performance since more options are available in the tables, requiring more bits to encode strings. The default value is false, meaning that buffers, string tables, etc., are cleared between each stanza. (This option is EXI/XMPP specific.)

The transmission of EXI-compressed stanzas takes the form of a sequence of EXI bodies. In order for the recipient to be able to correctly interpret these incoming EXI bodies, the sender is required to flush any pending bits at the end of the last End Document (ED) event for each stanza and then send any pending bytes available in the output buffer. Since this makes sure each EXI body starts at an even byte boundary, it permits the recipient to decompress the body into an XML stanza.

Therefore, each stanza sent on the stream, must be compressed separately, reusing the same options as used by the stream. (Options are not sent on the stream, only the generated EXI bodies).

Compression of the stanza must be done in document mode, not fragment mode, including the Start Document (SD) and End Document (ED) events. If there are unwritten bits pending after the last End Document (ED) event (after the end of the stanza), Zero-bits are written until a byte boundary is created. The receptor must ignore bits in the last byte after the last End Document event has been received.

During setup of the EXI compression engine, the client can choose if buffers are to be reused between stanzas, or cleared between each stanza. This is done using the EXI over XMPP specific option sessionWideBuffers, which is false by default, meaning buffers and string tables are cleared between each stanza.

There may be cases where maintaining buffers and string tables throughout the session is preferrable. Since strings are already available in the buffers, they don't need to be output in the stream the first time they appear in a stanza. However, the number of strings in tables increase, and so does the number of bits required to encode them. Depending on what type of communication is performed, this option might give better results one way or another. If the same type of message is always sent, maintaining string buffers may be more efficient. But if the client sends many many different types of messages, clearing buffers may be more efficient.

Note that the stream of EXI bodies is indefinite. It only stops when the session is closed, i.e. when the socket connection is dropped. Therefore, the buffers can grow indefinitely unless control is maintained on what types of messages are sent, their contents (specifically string values), and to whom they are sent (JIDs being strings). All string tables and buffers must be cleared when a connection is lost.

Note also that if you want the option to enter a session in the middle of the flow to listen to the communication, you need to clear tables and buffers between each stanza, or you will not be able to decode the binary stream appropriately.

Normally, prefixes are not preserved during EXI compression and decompression. If the communicating parties (sending client, XMPP server(s) and receiving clients) interpret incoming stanzas and content according to namespace, this should be sufficient. However, some implementations do not check namespaces, but prefix names used. In such cases, all communicating parties are required to enable the preserve prefixes option during negotionating.

Note: It is not sufficient that one party enable this option. Both sender and received are required to enable this option, or prefix names will be lost in the transmission.

Note also, that preserving prefix names result in less efficient compression. Therefore, all clients implementing EXI compression should strive to parse incoming XML based on namespace, not prefix name.

To successfully implement a network with clients having limited memory, such as sensor networks, care should be taken to make sure necessary schema files are preinstalled on the server, to avoid the necessity to upload schema files from the clients. Clients with limited memory might be unable to perform this task.

An alternative may be to install a richer client, that can upload the schema files to the server dynamically, and installing it into the network. Any client uploading a schema file, will make that schema file available for EXI compression to any other client in the network.

Schema files uploaded to the server should be cached on the server in some kind of schema repository. If memory is limited on the server, schema files should be sorted by last access. Schema files with the oldest last access timestamp could be removed to maintain the cache within an approved cache size.

Note that schema files have three keys: Target namespace, byte size and MD5 Hash. Multiple versions of a schema file may exist (that is, with the same target namespace but different byte sizes or MD5 hash codes). Note also, that for any practical purpose, schema files can be stored using only the MD5 hash as a key, since it is highly improbable that two different schema files will have the same MD5 hash (unless consciously created that way). MD5 hash values are always in lower case.

When the server lacks information about a given XML schema, the client has two options for updating the server. Either it uploads the schema, or it asks the server to download one.

Uploading a schema has the advantage, that the client knows exactly the version that the server requires. It has the disadvantage, that the client needs to store the schema and send a possible large schema to the server. If EXI is used because the device has limited memory, uploading a schema might not be an option.

Downloading a schema has the advantage, that size of schema does not matter. The disadvantage is that asynchronous errors might occur, so the client needs to pay attention to the responses returned by the server when downloading schemas. Also, downloading a schema, might download a version which does not correspond to the desired version of the schema. So, it's more important in this case that the client checks that the server actually has the version of the schema required by the client.

If two XMPP clients communicate with each other through an XMPP server, and both clients use EXI compression, the server must only forward binary packets if both EXI compressed channels have exactly the same setup. If any parameter is different, the server MUST always recompress packets sent through it.

Since the server always needs to decompress incoming EXI compressed packets to decode headers, omitting the compression part might save the server some processing power, but not all. Note that, in some networks it might be common using similar compression settings, while in others different compression settings are most common.

Also note that binary forwarding is only possible if session-wide buffers are not used.

Errata and other updates may happen to well-known schemas. Slightest modification to XML schemas may break interoperability of EXI nodes. However, negotiating everything is not efficient. Well-known and aged schemas that referred from the schema for EXI/XMPP shall be snapshoted for use in EXI processors.

Actual snapshot is located in (TBD-URL).

The interface between the XMPP engine, whether in the client or the server, and the EXI compression engine is required to provide the engine with ONE XML Schema to use during the compression. The Efficient XML Interchange (EXI) Format also specifies the use of Schema IDs identifying the schema to use.

However, in the XMPP case, the schema to provide to the EXI compression engine must be created dynamically based on the handshake provided during setup of the connection. Since this generation must be done both on the server side as well as the client side, it is important that the schemas be created semantically equivalent. This section describes how to create such schemas, henceforth called Canonical Schemas is described in this section.

A canonical schema is simply a wrapper importing each of the schemas negotiated for the connection. The schemas MUST be imported in ascending namespace order.

After generating the canonical schema, it's a good idea to create a corresponding Configuration ID. The Configuration ID however, includes not only the schemas imported into the canonical schema, but also the EXI options to use during compression/decompression. The canonical schema should be persisted for simple reuse when quick setup is used.

The target namespace of the canonical schema MUST be urn:xmpp:exi:cs.

The Schema ID to use is irrelevant in the XMPP layer of communiction. Therefore, the server and client can create their own Schema IDs, according to some algorithm. It is not important if the Schema IDs match, since they are not used in data transmitted between the client and server.

The configurationId attribute has a similar attribute called configurationLocation. This attribute provides a mechanism to setup an EXI connection rapidly using option documents installed as files on the server or available on the network accessible through an URL.

A client may specify a configurationId or configurationLocation on exi:setup element. If the server has corresponding setup configuration, the server may respond with an exi:setupResponse with agreement="true". If the server does not know the configurationId or does not be able to use the given configurationLocation, the server shall respond with an exi:setupResponse with agreement="false".

This specification does not define the format of this Configuration Location attribute, and so it is server specific, or if it is supported. If used on a server not supporting this attribute, or if the contents of the attribute is invalid, the server returns an agreement=false response. Otherwise the semantics of the configurationLocation attribute is the same as for the configurationId attribute, except it provide a mechism for static configurations, while the configurationId provides a mechanism for dynamic configurations.

The format for these opton documents or locations is beyond the scope of this specification.

The format for these opton documents or locations is beyond the scope of this specification.

The following table lists XMPP schemas at the time of writing, and their corresponding bytes sizes and MD5 Hash values.

The following sections list some known problems that might affect already defined schemas. For these schemas to be used together with EXI compression, the recommended procedures should be considered. If changes to the schema is required, new byte sizes and hash values must be computed for the changed schema correspondingly.

Patched version of XML schemas may be placed in the snapshot repository for well-known schemas to avoid wild patches (and too many derived schema for the same model).

Default EXI grammar used in EXI/XMPP SHALL be equivalent to the EXI grammar defined by the following schema. The default EXI grammar is used in initiating connection of EXI/XMPP alternate binding. The default schema is defined by following definition, and all the imported schemas SHOULD be same schemas described in the snapshot and SHOULD have the identical MD5 hash value described in this section.

The ${snapshot-url} corresponds to yet-to-be-specified schema snapshot repository described in snapshot repository for well-known schemas section. The schemaId of this schema will is 'urn:xmpp:exi:default'. SchemaId is ID for this instance and not a namespace identifier and intentionally different from the target namespace of the schema.

Following schema Ids are reserved.

• Schema IDs starts with "c:". This is used as shortcut setup for alternative transport binding.
• Schema IDs starts with "urn:xmpp". This may be used to describe XEP-based schemas.