<?xml version="1.0" encoding="latin1" ?>
<!DOCTYPE chapter SYSTEM "chapter.dtd">
<chapter>
<header>
<copyright>
<year>1997</year><year>2009</year>
<holder>Ericsson AB. All Rights Reserved.</holder>
</copyright>
<legalnotice>
The contents of this file are subject to the Erlang Public License,
Version 1.1, (the "License"); you may not use this file except in
compliance with the License. You should have received a copy of the
Erlang Public License along with this software. If not, it can be
retrieved online at http://www.erlang.org/.
Software distributed under the License is distributed on an "AS IS"
basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
the License for the specific language governing rights and limitations
under the License.
</legalnotice>
<title>Advanced Agent Topics</title>
<prepared></prepared>
<responsible></responsible>
<docno></docno>
<approved></approved>
<checked></checked>
<date></date>
<rev></rev>
<file>snmp_advanced_agent.xml</file>
</header>
<p>The chapter <em>Advanced Agent Topics</em> describes the more advanced
agent related features of the SNMP development tool. The following topics
are covered:
</p>
<list type="bulleted">
<item>When to use a Sub-agent</item>
<item>Agent semantics</item>
<item>Sub-agents and dependencies</item>
<item>Distributed tables</item>
<item>Fault tolerance</item>
<item>Using Mnesia tables as SNMP tables</item>
<item>Audit Trail Logging</item>
<item>Deviations from the standard
</item>
</list>
<section>
<title>When to use a Sub-agent</title>
<p>The section <em>When to use a Sub-agent</em> describes situations
where the mechanism of loading and unloading MIBs is insufficient.
In these cases a sub-agent is needed.
</p>
<section>
<title>Special Set Transaction Mechanism</title>
<p>Each sub-agent can implement its own mechanisms for
<c>set</c>, <c>get</c> and <c>get-next</c>. For example, if the
application requires the <c>get</c> mechanism to be
asynchronous, or needs a N-phase <c>set</c> mechanism, a
specialized sub-agent should be used.
</p>
<p>The toolkit allows different kinds of sub-agents at the same
time. Accordingly, different MIBs can have different <c>set</c>
or <c>get</c> mechanisms.
</p>
</section>
<section>
<title>Process Communication</title>
<p>A simple distributed agent can be managed without sub-agents.
The instrumentation functions can use distributed Erlang to
communicate with other parts of the application. However, a
sub-agent can be used on each node if this generates too much
unnecessary traffic. A sub-agent processes requests per
incoming SNMP request, not per variable. Therefore the network
traffic is minimized.
</p>
<p>If the instrumentation functions communicate with UNIX
processes, it might be a good idea to use a special
sub-agent. This sub-agent sends the SNMP request to the other
process in one packet in order to minimize context switches. For
example, if a whole MIB is implemented on the C level in UNIX,
but you still want to use the Erlang SNMP tool, then you may
have one special sub-agent, which sends the variables in the
request as a single operation down to C.
</p>
</section>
<section>
<title>Frequent Loading of MIBs</title>
<p>Loading and unloading of MIBs are quite cheap
operations. However, if the application does this very often,
perhaps several times per minute, it should load the MIBs once
and for all in a sub-agent. This sub-agent only registers and
unregisters itself under another agent instead of loading the
MIBs each time. This is cheaper than loading an MIB.
</p>
</section>
<section>
<title>Interaction With Other SNMP Agent Toolkits</title>
<p>If the SNMP agent needs to interact with sub-agents
constructed in another package, a special sub-agent should be
used, which communicates through a protocol specified by the
other package.
</p>
</section>
</section>
<section>
<title>Agent Semantics</title>
<p>The agent can be configured to be multi-threaded, to process
one incoming request at a time, or to have a request limit
enabled (this can be used for load control or to limit the effect
of DoS attacks). If it is multi-threaded, read requests (<c>get</c>,
<c>get-next</c> and <c>get-bulk</c>) and traps are processed in
parallel with each other and <c>set</c> requests. However, all
<c>set</c> requests are serialized, which means that if the agent
is waiting for the application to complete a complicated write
operation, it will not process any new write requests until this
operation is finished. It processes read requests and sends traps,
concurrently. The reason for not handle write requests in parallel is
that a complex locking mechanism would be needed even in the simplest
cases. Even with the scheme described above, the user must be
careful not to violate that the <c>set</c> requests are atoms.
If this is hard to do, do not use the multi-threaded feature.
</p>
<p>The order within an request is undefined and variables are not
processed in a defined order. Do not assume that the first
variable in the PDU will be processed before the second, even if
the agent processes variables in this order. It
cannot even be assumed that requests belonging to different
sub-agents have any order.
</p>
<p>If the manager tries to set the same variable many times in the
same PDU, the agent is free to improvise. There is no definition
which determines if the instrumentation will be called once or
twice. If called once only, there is no definition that determines
which of the new values is going to be supplied.
</p>
<p>When the agent receives a request, it keeps the request ID for
one second after the response is sent. If the agent receives
another request with the same request ID during this time, from
the same IP address and UDP port, that request will be
discarded. This mechanism has nothing to do with the function
<c>snmpa:current_request_id/0</c>.</p>
</section>
<section>
<title>Sub-agents and Dependencies </title>
<p>The toolkit supports the use of different types of sub-agents,
but not the construction of sub-agents.
</p>
<p>Also, the toolkit does not support dependencies between
sub-agents. A sub-agent should by definition be stand alone and it is
therefore not good design to create dependencies between them.
</p>
</section>
<section>
<title>Distributed Tables</title>
<p>A common situation in more complex systems is that the data in
a table is distributed. Different table rows are implemented in
different places. Some SNMP tool-kits dedicate an SNMP sub-agent for
each part of the table and load the corresponding MIB into all
sub-agents. The Master Agent is responsible for presenting the
distributed table as a single table to the manager. The toolkit
supplied uses a different method.
</p>
<p>The method used to implement distributed tables with this SNMP
tool is to implement a table coordinator process responsible for
coordinating the processes, which hold the table data and they
are called table holders. All table holders must in some way be
known by the coordinator; the structure of the table data
determines how this is achieved. The coordinator may require
that the table holders explicitly register themselves and specify
their information. In other cases, the table holders can be
determined once at compile time.
</p>
<p>When the instrumentation function for the distributed table is
called, the request should be forwarded to the table
coordinator. The coordinator finds the requested information among
the table holders and then returns the answer to the
instrumentation function. The SNMP toolkit contains no support for
coordination of tables since this must be independent of the
implementation.
</p>
<p>The advantages of separating the table coordinator from the
SNMP tool are:
</p>
<list type="bulleted">
<item>We do not need a sub-agent for each table holder. Normally,
the sub-agent is needed to take care of communication, but in
Distributed Erlang we use ordinary message passing.
</item>
<item>Most likely, some type of table coordinator already
exists. This process should take care of the instrumentation for
the table.
</item>
<item>The method used to present a distributed table is strongly
application dependent. The use of different masking techniques
is only valid for a small subset of problems and registering
every row in a distributed table makes it non-distributed.
</item>
</list>
</section>
<section>
<title>Fault Tolerance</title>
<p>The SNMP agent toolkit gets input from three different sources:
</p>
<list type="bulleted">
<item>UDP packets from the network</item>
<item>return values from the user defined instrumentation functions</item>
<item>return values from the MIB.
</item>
</list>
<p>The agent is highly fault tolerant. If the manager gets an
unexpected response from the agent, it is possible that some
instrumentation function has returned an erroneous value. The
agent will not crash even if the instrumentation does. It should
be noted that if an instrumentation function enters an infinite
loop, the agent will also be blocked forever. The supervisor ,or
the application, specifies how to restart the agent.
</p>
<section>
<title>Using the SNMP Agent in a Distributed Environment</title>
<p>The normal way to use the agent in a distributed
environment is to use one master agent located at one node,
and zero or more sub-agents located on other nodes. However,
this configuration makes the master agent node a single point
of failure. If that node goes down, the agent will not work.
</p>
<p>One solution to this problem is to make the snmp application
a distributed Erlang application, and that means, the agent
may be configured to run on one of several nodes. If the node
where it runs goes down, another node restarts the agent.
This is called <em>failover</em>. When the node starts again,
it may <em>takeover</em> the application. This solution to
the problem adds another problem. Generally, the new node has
another IP address than the first one, which may cause
problems in the communication between the SNMP managers and
the agent.
</p>
<p>If the snmp agent is configured as a distributed Erlang
application, it will during takeover try to load the same MIBs
that were loaded at the old node. It uses the same filenames
as the old node. If the MIBs are not located in the same
paths at the different nodes, the MIBs must be loaded
explicitly after takeover.
</p>
</section>
</section>
<section>
<title>Using Mnesia Tables as SNMP Tables</title>
<p>The Mnesia DBMS can be used for storing data of SNMP
tables. This means that an SNMP table can be implemented as a
Mnesia table, and that a Mnesia table can be made visible via
SNMP. This mapping is largely automated.
</p>
<p>There are three main reasons for using this mapping:
</p>
<list type="bulleted">
<item>We get all features of Mnesia, such as fault tolerance,
persistent data storage, replication, and so on.
</item>
<item>Much of the work involved is automated. This includes
<c>get-next</c> processing and <c>RowStatus</c> handling.
</item>
<item>The table may be used as an ordinary Mnesia table, using
the Mnesia API internally in the application at the same time as
it is visible through SNMP.
</item>
</list>
<p>When this mapping is used, insertion and deletion in the
original Mnesia table is slower, with a factor O(log n). The read
access is not affected.
</p>
<p>A drawback with implementing an SNMP table as a Mnesia table is
that the internal resource is forced to use the table definition
from the MIB, which means that the external data model must be
used internally. Actually, this is only partially true. The Mnesia
table may extend the SNMP table, which means that the Mnesia table
may have columns which are use internally and are not seen by
SNMP. Still, the data model from SNMP must be maintained. Although
this is undesirable, it is a pragmatic compromise in many
situations where simple and efficient implementation is preferable
to abstraction.
</p>
<section>
<title>Creating the Mnesia Table</title>
<p>The table must be created in Mnesia before the manager can
use it. The table must be declared as type <c>snmp</c>. This
makes the table ordered in accordance with the lexicographical
ordering rules of SNMP. The name of the Mnesia table must be
identical to the SNMP table name. The types of the INDEX fields
in the corresponding SNMP table must be specified.
</p>
<p>If the SNMP table has more than one INDEX column, the
corresponding Mnesia row is a tuple, where the first element
is a tuple with the INDEX columns. Generally, if the SNMP table
has <em>N</em> INDEX columns and <em>C</em> data columns, the
Mnesia table is of arity <em>(C-N)+1</em>, where the key is a
tuple of arity <em>N</em> if <em>N > 1</em>, or a single term
if <em>N = 1</em>.
</p>
<p>Refer to the Mnesia User's Guide for information on how to
declare a Mnesia table as an SNMP table.
</p>
<p>The following example illustrates a situation in which we
have an SNMP table that we wish to implement as a Mnesia
table. The table stores information about employees at a
company. Each employee is indexed with the department number and
the name.
</p>
<code type="none">
empTable OBJECT-TYPE
SYNTAX SEQUENCE OF EmpEntry
ACCESS not-accessible
STATUS mandatory
DESCRIPTION
"A table with information about employees."
::= { emp 1}
empEntry OBJECT-TYPE
SYNTAX EmpEntry
ACCESS not-accessible
STATUS mandatory
DESCRIPTION
""
INDEX { empDepNo, empName }
::= { empTable 1 }
EmpEntry ::=
SEQUENCE {
empDepNo INTEGER,
empName DisplayString,
empTelNo DisplayString
empStatus RowStatus
}
</code>
<p>The corresponding Mnesia table is specified as follows:
</p>
<code type="none">
mnesia:create_table([{name, employees},
{snmp, [{key, {integer, string}}]},
{attributes, [key, telno, row_status]}]).
</code>
<note>
<p>In the Mnesia tables, the two key columns are stored as a
tuple with two elements. Therefore, the arity of the table is
3.</p>
</note>
</section>
<section>
<title>Instrumentation Functions</title>
<p>The MIB table shown in the previous section can be compiled
as follows:
</p>
<pre>
1> <input>snmpc:compile("EmpMIB", [{db, mnesia}]).</input>
</pre>
<p>This is all that has to be done! Now the manager can read,
add, and modify rows. Also, you can use the ordinary Mnesia API
to access the table from your programs. The only explicit action
is to create the Mnesia table, an action the user has to perform
in order to create the required table schemas.</p>
</section>
<section>
<title>Adding Own Actions</title>
<p>It is often necessary to take some specific action when a
table is modified. This is accomplished with an instrumentation
function. It executes some specific code when the table is set,
and passes all other requests down to the pre-defined function.
</p>
<p>The following example illustrates this idea:
</p>
<code type="none">
emp_table(set, RowIndex, Cols) ->
notify_internal_resources(RowIndex, Cols),
snmp_generic:table_func(set, RowIndex, Cols, {empTable, mnesia});
emp_table(Op, RowIndex, Cols) ->
snmp_generic:table_func(Op, RowIndex, Cols, {empTable, mnesia}).
</code>
<p>The default instrumentation functions are defined in the
module <c>snmp_generic</c>. Refer to the Reference Manual,
section SNMP, module <c>snmp_generic</c> for details.</p>
</section>
<section>
<title>Extending the Mnesia Table</title>
<p>A table may contain columns that are used internally, but
should not be visible to a manager. These internal columns must
be the last columns in the table. The <c>set</c> operation will
not work with this arrangement, because there are columns that
the agent does not know about. This situation is handled by
adding values for the internal columns in the <c>set</c>
function.
</p>
<p>To illustrate this, suppose we extend our Mnesia
<c>empTable</c> with one internal column. We create it as
before, but with an arity of 4, by adding another attribute.
</p>
<code type="none">
mnesia:create_table([{name, employees},
{snmp, [{key, {integer, string}}]},
{attributes, {key, telno, row_status, internal_col}}]).
</code>
<p>The last column is the internal column. When performing a
<c>set</c> operation, which creates a row, we must give a
value to the internal column. The instrumentation functions will now
look as follows:
</p>
<code type="none">
-define(createAndGo, 4).
-define(createAndWait, 5).
emp_table(set, RowIndex, Cols) ->
notify_internal_resources(RowIndex, Cols),
NewCols =
case is_row_created(empTable, Cols) of
true -> Cols ++ [{4, "internal"}]; % add internal column
false -> Cols % keep original cols
end,
snmp_generic:table_func(set, RowIndex, NewCols, {empTable, mnesia});
emp_table(Op, RowIndex, Cols) ->
snmp_generic:table_func(Op, RowIndex, Cols, {empTable, mnesia}).
is_row_created(Name, Cols) ->
case snmp_generic:get_status_col(Name, Cols) of
{ok, ?createAndGo} -> true;
{ok, ?createAndWait} -> true;
_ -> false
end.
</code>
<p>If a row is created, we always set the internal column to
<c>"internal"</c>.
</p>
</section>
</section>
<section>
<title>Deviations from the Standard</title>
<p>In some aspects the agent does not implement SNMP fully. Here
are the differences:
</p>
<list type="bulleted">
<item>The default functions and <c>snmp_generic</c> cannot
handle an object of type <c>NetworkAddress</c> as INDEX
(SNMPv1 only!). Use <c>IpAddress</c> instead.
</item>
<item>The agent does not check complex ranges specified for
INTEGER objects. In these cases it just checks that the value
lies within the minimum and maximum values specified. For
example, if the range is specified as <c>1..10 | 12..20</c>
the agent would let 11 through, but not 0 or 21. The
instrumentation functions must check the complex ranges
itself.
</item>
<item>The agent will never generate the <c>wrongEncoding</c>
error. If a variable binding is erroneous encoded, the
<c>asn1ParseError</c> counter will be incremented.
</item>
<item>A <c>tooBig</c> error in an SNMPv1 packet will always use
the <c>'NULL'</c> value in all variable bindings.
</item>
<item>The default functions and <c>snmp_generic</c> do not check
the range of each OCTET in textual conventions derived from
OCTET STRING, e.g. <c>DisplayString</c> and
<c>DateAndTime</c>. This must be checked in an overloaded
<c>is_set_ok</c> function.
</item>
</list>
</section>
</chapter>
