Design considerations
The PPI interface is designed to be as simple as possible. It provides sane defaults for all
configurable parameters to simplify getting started. It also automates all resource
management. The throttling infrastructure handles blocking conditions (like input exhaustion)
automatically.
Packets
The PPI processes packets and uses the <a
href="/api/trunk/senf/Packets/doc/html/index.html">Packet library</a> to handle them. All packets
are internally passed around as generic \ref senf::Packet references, however connectors may
optionally be defined as sending or receiving packets of a specific type only.
Modules
A module is represented by a class derived from senf::ppi::module::Module. Each module has
several components:
\li It may have any number of \ref ppi_connectors (inputs and outputs)
\li Each module declares flow information which details the route packets take within the
module. This information does not define how the information is processed, it only tells,
where data arriving on some input will be directed at (\ref
senf::ppi::module::Module::route())
\li The module might take additional parameters.
\li The module might also register additional \ref ppi_events.
Generally, modules are divided into several categories:
\li \ref io_modules receive external data or forward packets out of the PPI
\li \ref routing_modules forward packets within the network
\li \ref adapter_modules are used to connect incompatible connectors to each other
\li Application modules are modules implemented to perform an applications function
Of these modules, normally only the application modules need to be implemented since the library
provides an extensive set of reusable modules.
The following example module declares three \ref ppi_connectors "Connectors": \c payload,
\c stuffing and \c output. These connectors are defined as \e public data members so they
can be accessed from the outside. This is important as we will see below.
{
public:
: timer_(1000u, packetsPerSecond)
{
route(payload, output);
route(stuffing, output);
registerEvent( timer_, &RateStuffer::tick );
}
private:
void tick()
{
if (payload)
else
}
};
The constructor will declare flow information using senf::ppi::module::Module::route(). Then the
module registers an interval timer which will fire <tt>packetsPerSecond</tt> times every
<tt>1000</tt> milliseconds.
The module processing is very simple: Whenever a timer tick arrives a packet is sent. If the \c
payload input is ready (see \ref ppi_throttling), a payload packet is sent, otherwise a stuffing
packet is sent. The module will therefore provide a constant stream of packets at a fixed rate
on \c output (see the
<a href="/api/trunk/Examples/RateStuffer/doc/html/index.html">RateStuffer</a> example application
for a slightly different approach)
An example module to generate the stuffing packets could be
class CopyPacketGenerator
{
public:
CopyPacketGenerator(Packet template)
: template_ (template)
{
output.
onRequest(&CopyPacketGenerator::makePacket);
}
private:
Packet template_;
void makePacket()
{
output(template_.clone());
}
};
This module just produces a copy of a given packet whenever output is requested.
\see senf::ppi::module::Module
Connectors
The input and output attachment points of a module are called connectors. Each connector may be
active or passive. This gives us 4 types of connectors:
\li senf::ppi::connector::ActiveInput
\li senf::ppi::connector::ActiveOutput
\li senf::ppi::connector::PassiveInput
\li senf::ppi::connector::PassiveOutput
An \e active connector (input or output) is <em>activated by the module</em> to send data or to
poll for available packets. This means, the modules processing routine will call the connector
without being signaled by the framework to read the connector. It just actively fetches a
Packet.
A \e passive connector is <em>signaled by the framework</em> to fetch data from the module or to
pass data into the module. The module must register a callback which will be called, whenever a
packet is requested from the module or whenever a new packet is made available for the module to
process.
To send or receive a packet (either actively or passively) the module just calls the
connector. It is permissible to generate or process multiple packets in one iteration. However,
you must ensure yourself that enough packets are available to be read if more than one packet
shall be read. It is also permissible to not handle a packet at all even if signaled to do
so. The packet will automatically be queued.
To provide this flexibility, all input connectors incorporate a packet queue. This queue is
exposed to the module and allows the module to optionally process packets in batches.
Connectors take an optional template argument which allows to specify the type of packet this
connector sends or received. This template arguments defaults to \ref senf::Packet.
\see \ref senf::ppi::connector
Connections
\image html ratestuffer.png Simple RateStuffer
To make use of the modules, they have to be instantiated and connections have to be created
between its connectors. It is possible to connect any pair of input/output connectors as long as
one of them is active and the other is passive.
It is possible to connect two active or passive connectors with each other using a special
adaptor module (senf::ppi::module::PassiveQueue or senf::ppi::module::ActiveFeeder
respectively).
Additionally, the connectors must be type-compatible: Either one (or both) of the connectors
must be untyped (they accept arbitrary senf::Packet's, the optional template argument is empty),
or they both accept the same type of packet. This check is performed at runtime.
To complete our simplified example: Lets connect senf::ppi::module::ActiveSocketReader and
senf::ppi::module::PassiveSocketWriter to our example module:
senf::Packet stuffingPacket = senf::DataPacket::create(...);
CopyPacketGenerator generator (stuffingPacket);
adaptor.qdisc(ThresholdQueueing(10,5));
This application will read udp-packets coming in on port 1111 and will forward
them to port 2222 on host 2.3.4.5 with a fixed rate of 10 packets / second.
We start out by instantiating the necessary modules. Then the connections between these modules
are set up by successively connecting each output connector to an input connector. As can be
seen, the name of the connector can be left of if it is named \c output or \c input
respectively.
The buffering on the udpInput <-> rateStuffer adaptor is changed so the queue will begin to
throttle only if more than 10 packets are in the queue. The connection will be unthrottled as
soon as there are no more than 5 packets left in the queue (see \ref ppi_throttling).
Throttling
Throttling and throttle notifications at it's base is about handling blocking conditions. The
most straight forward blocking condition is that of a file descriptor not being available for
reading resp. writing. Other blocking conditions can arise for example when a queue fills up or
if a module for some application specific reason does not want to handle packets for a period of
time.
All this is handled using throttle notifications. We need throttle notifications so a passive
connector can tell it's connected peer that it cannot service further requests until an
unthrottle notification is sent. This tells us, that from the view of someone implementing a
module, throttle notifications will always be received on active connectors and be sent on
passive connectors.
This tells us, that the direction of control flow (the throttle notifications) is from passive
to active connectors and does \e not depend on the direction of data flow (which flows from
output to input connector). Thinking about this, this makes sense: The module with the active
connector is the one initiating the data processing (after all, it is the \e active part) and
needs to be told not to request or send packets on it's connector since the connected passive
peer cannot handle the request.
So if a passive connector cannot handle requests, the connector must be \e throttled. Throttling
the connector will forward a throttle notification to its peer. The peer then handles the
throttling notification.
There are two ways, throttle notifications can be handled: By automatic throttling or by
registering callbacks. The default is <em>automatic throttling</em>.
<em>Automatic throttling</em> is based on the routing information available to the module. Every
notification received is forwarded within the module along all known routes from active to
passive connectors (routes which connect to active or passive connectors are absolutely valid,
they just are not \e forwarding routes, they are ignored by the throttle
notifications). Together with automatic event throttling (see \ref ppi_events), this is all that
is normally needed to handle throttle notifications: By forwarding the notifications we ensure,
that a module's passive connectors will only be signaled when it's corresponding active
connectors are not throttled (as defined by the routing information). The module is therefore
not called until the connector(s) are untrhottled.
<em>Throttle callbacks</em> can optionaly be registerd (with automatic throttling enabled or
disabled, see \ref senf::ppi::connector::ActiveConnector) to be called when a throttle
notification is received. The callback may then handle the notification however it sees fit, for
example by manually throttling some passive connector (see \ref
senf::ppi::connector::PassiveConnector).
To enable/disable automatic throttling, the \ref senf::ppi::module::Module::route() command
returns a reference to a \ref senf::ppi::Route instance. If this route is \e forwarding route,
(that is, of the connectors is passive and the other is active), the return value will be
derived from \ref senf::ppi::ForwardingRoute which provides members to control the throttle
notification forwarding.
\see
senf::ppi::module::Module \n
senf::ppi::Route
Events
Modules may register additional events. These external events are very important since they
drive the PPI framework. Events are like external calls into the module network which are sent
whenever some event happens. Some possible events are
\li timer events (senf::ppi::IntervalTimer)
\li read or write events on some file descriptor (senf::ppi::IOEvent)
\li internal events (senf::ppi::IdleEvent)
The PPI really is not concerned, how the events are called and what information is needed to
perform the call. This is handled by the <a
href="/api/trunk/senf/Scheduler/doc/html/index.html">Scheduler</a>, which is wrapped by the event
classes.
All events are derived from senf::ppi::EventDescriptor. The base class allows to enable and
disable the event. Each type of event will take descriptor specific constructor arguments to
describe the event to be generated. Events are declared as (private) data members of the
module and are then registered using senf::ppi::module::Module::registerEvent().
Each event when signaled is described by an instance of the descriptor specific \e
descriptorType \c ::Event class. This instance will hold the event specific information (like
scheduled time of the event, file handle state and so on). This information is passed to the
callback.
Additionally, events are valid routing targets. This feature allows events to be disabled and
enabled by throtling notifications. For the sake of routing, an event may be used like an active
input or output. Iit is \e active from the PPI's point of view since it is signaled from the
outside and not by some module. It may be either input or output depending on the operation the
event controls.
If we take into account event routing, we can extend the \c RateStuffer constructor accordingly:
: timer_(1000u, packetsPerSecond)
{
route(payload, output);
route(stuffing, output);
route(timer_, output);
registerEvent( timer_, &RateStuffer::tick );
}
We have added the marked route call. This way, the \c timer_ will receive throttling
notifications from the output: Whenever the output is throttled, the event will be disabled
until the output is unthrottled again.
\see senf::ppi::EventDescriptor
Running the network
After the network has been set up, senf::ppi::run() is called to execute it. This call will only
return after all data has been processed. The PPI knows this, when no events are enabled any
more. Without events, nothing will happen any more since it is the events which drive the
PPI. Therefore the PPI surmises, that all data has been processed and returns from
senf::ppi::run().
This works very well with automatic throttling. When no data is available to be processed any
more and no more data can be expected to arrive (for Example since data has been read from a
file which is now exhausted) all events will be disabled automatically via trhottle
notifications and so signal that any processing should stop.
Information Flow
The above description conceptually introduces three different flow levels:
\li The <em>data flow</em> is, where the packets are flowing. This flow always goes from output
to input connector.
\li The <em>execution flow</em> describes the flow of execution from one module to another. This
flow always proceeds from active to passive connector.
\li The <em>control flow</em> is the flow of throttling notifications. This flow always proceeds
\e opposite to the execution flow, from passive to active connector.
This is the outside view, from without any module. These flows are set up using
senf::ppi::connect() statements.
Within a module, the different flow levels are defined differently depending on the type of
flow:
\li The <em>data flow</em> is defined by how data is processed. The different event and
connector callbacks will pass packets around and thereby define the data flow
\li Likewise, the <em>execution flow</em> is defined parallel to the data flow (however possible
in opposite direction) by how the handler of one connector calls other connectors.
\li The <em>control flow</em> is set up using senf::ppi::Module::route statements (as long as
automatic throttling is used. Manual throttling defines the control flow within the
respective callbacks).
In nearly all cases, these flows will be parallel. Therefore it makes sense to define the \c
route statement as defining the 'conceptual data flow' since this is also how control messages
should flow (sans the direction, which is defined by the connectors active/passive property).
\see \ref ppi_implementation