+\section{Discussion \& Future Research}
+The system is still a crude prototype and a proof of concept. Improvements and
+extension for the system are amply available in several fields of study.
+
+\subsection{Simulation}
+An additional simulation view to the \gls{mTask}-\gls{EDSL} could be added that
+works in the same way as the existing \gls{C}-backed simulation. It simulates
+the bytecode interpretation. Moreover, it would be possible to let the simulator
+function as a real device, thus handling all communication through the
+existing \gls{SDS}-based systems and behave like a real device. At the moment
+the \emph{POSIX}-client is the reference client and contains debugging code.
+Adding a simulation view to the system allows for easy interactive debugging.
+However, it might not be easy to devise a simulation tool that accurately
+simulates the \gls{mTask} system on some levels. The semantics can be simulated
+but timing and peripheral input/output are more difficult to simulate properly.
+
+\subsection{Optimization}
+\paragraph{Multitasking on the client:}
+True multitasking could be added to the client software. This allows
+\gls{mTask}-\glspl{Task} to run truly parallel. All \glspl{mTask} get slices
+of execution time and will each have their own interpreter state instead of a
+single system-wide state which is reset after am \gls{mTask} finishes. This
+does require separate stacks for each \gls{Task} and therefore increases the
+system requirements of the client software. However, it could be implemented as
+a compile-time option and exchanged during the handshake so that the server
+knows the multithreading capabilities of the client.
+
+\paragraph{Optimizing the interpreter:}
+Due to time constraints and focus, hardly any work has been done in the
+interpreter. The current interpreter is a no nonsense stack machine. A lot of
+improvements can be done in this part. For example, precomputed \emph{gotos}
+can improve jumping to the correct part of the code corresponding to the
+correct instruction. Moreover, the stack currently consists of 16-bit values.
+All operations work on 16-bit values and this simplifies the interpreter
+implementation. A memory improvement can be made by converting the stack to
+8-bit values. This does pose some problems since an equality instruction must
+work on single-byte booleans \emph{and} two-byte integers. Adding specialized
+instructions per word size could overcome this problem.
+
+\subsection{Resources}
+\paragraph{Resource analysis: }
+Resource analysis during compilation can be useful to determine if an
+\gls{mTask}-\gls{Task} is suitable for a specific device. If the device does
+not contain the correct peripherals --- such as an \gls{LCD} --- then the
+\gls{mTask}-\gls{Task} should be rejected and feedback to the user must be
+given. It might even be possible to do this statically on the type level.
+
+\paragraph{Extended resource analysis: }
+The previous idea could be extended to the analysis of stack size and possibly
+communication bandwidth. With this functionality ever more reliable fail-over
+systems can be designed. When the system knows precise bounds it can allocate
+more \glspl{Task} on a device whilst staying within safe memory bounds. The
+resource allocation can be done at runtime within the backend itself or a
+general backend can be devised that can calculate the resources needed for a
+given \gls{mTask}. A specific \gls{mTask} cannot have multiple views at the
+same time due to the restrictions of class based shallow embedding. It might
+even be possible to encode the resource allocation in the type system itself
+using forms of dependant types.
+
+\subsection{Functionality}
+\paragraph{Add more combinators: }
+More \gls{Task}-combinators --- already existing in the \gls{iTasks}-system ---
+could be added to the \gls{mTask}-system to allow for more fine-grained control
+flow between \gls{mTask}-\glspl{Task}. In this way the new system follows the
+\gls{TOP} paradigm even more and makes programming \glspl{mTask} for
+\gls{TOP}-programmers more seamless. Some of the combinators require previously
+mentioned extension such as the parallel combinator. Others might be achieved
+using simple syntactic transformations.
+
+\paragraph{Launch \glspl{Task} from a \gls{Task}: }\label{par:tasklaunch}
+Currently the \gls{C}-view allows \glspl{Task} to launch other \glspl{Task}. In
+the current system this type of logic has to take place on the server side.
+Adding this functionality to the bytecode-view allows greater flexibility,
+easier programming and less communication resources. Adding these semantics
+requires modifications to the client software and extensions to the
+communication protocol since relations between \glspl{Task} also need to be
+encoded and communicated.
+
+The \gls{SDS} functionality in the current system is bare. There is no easy way
+of reusing an \gls{SDS} for another \gls{Task} on the same device or on another
+device. Such functionality can be implemented in a crude way by tying the
+\glspl{SDS} together in the \gls{iTasks} environment. However, this will result
+in a slow updating system. Functionality for reusing shares from a device
+should be added. This requires rethinking the storage because some typedness is
+lost when the \gls{SDS} is stored after compilation. A possibility would be to
+use runtime typing with \CI{Dynamic}s or the encoding technique currently used
+for \CI{BCValue}s. Using \glspl{SDS} for multiple \glspl{Task} within one
+device is solved when the previous point at paragraph~\ref{par:tasklaunch} is
+implemented.
+
+\subsection{Robustness}
+\paragraph{Reconnect with lost devices:}
+The robustness of the system can be greatly improved. Devices that lose
+connection are not well supported in the current system. The device will stop
+functioning and has to be emptied for a reconnect. \Glspl{Task} residing on a
+device that disconnected should be kept on the server to allow a swift
+reconnect and restoration of the \glspl{Task}. This holds the same for the
+client software. The client drops all existing \glspl{Task} on a shutdown
+request. An extra specialization of the shutdown could be added that drops the
+connection but keeps the \glspl{Task} in memory. During the downtime the
+\glspl{Task} can still be executed but publications need to be delayed. If the
+same server connects to the client the delayed publications can be sent
+anyways.
+
+\paragraph{Reverse \gls{Task} sending:}
+Moreover, devices could send their current \glspl{Task} back at the
+server to synchronize it. This allows interchanging servers without
+interrupting the client. Allowing the client to send \glspl{Task} to the server
+is something to handle with care because it can easily cause high bandwidth
+usage.
+