96
7 Designing Application-Specific Architectures
number of experiments as well as the length of the experiments) has been increased
until the proposed method did not generate an architecture within 60 min.
Using five operations (a reasonable number of basic operations), the method is
capable of determining architectures for up to
• 32 experiments with an average experiment length of 6.28 modules (the resulting
architecture consists of 30 nodes and 170 edges).
• an average experiment length of 13 modules for 6 experiments (the resulting
architecture consists of 41 nodes and 103 edges).
7.5.2 Comparison to the Ring Architecture
The ring constitutes the commonly used architecture so far. However, this architecture frequently leads to cases where the payload droplet has to traverse the
ring several times (resulting in a high connection depth) or has to be re-injected
by the MPU. Application-specific architectures as generated by the proposed
method overcome these problems. In order to demonstrate this, the proposed
method has been used to generate application-specific architectures for different
sets of experiments. The used experiments are generated out of sequencing graphs
from [102]. More precisely, the operations along each path through the graph are
considered as an experiment. For the generated architectures, both the number of
connections and the number of modules are used as optimization criteria.
Table 7.2 summarizes the obtained results. The first five columns provide the
name, the number of operations (|O|), the number of experiments (||), the
maximal number of considered instances for each module (Max. Inst.), and the
average length of the experiments (Avg. |φ|). Afterwards, the respective results
are reported if the experiments have been realized by means of a ring 3 (denoted
by R) or an application-specific architecture (denoted by G). More precisely, the
number of modules (|V |) and connections (|E|) specify the size of the resulting
architecture. The maximal connection depth (Max. Conn. Depth) and the average
connection depth (Avg. Conn. Depth) state how many connections the payload
droplet has to flow through at most/on average in order to conduct any of the
experiments. The maximal payload re-injections (Max. Re-inj.) and the average
payload re-injections (Avg. Re-inj.) state how often the payload droplet needs
to be re-injected by the MPU at most/on average per experiment. Furthermore,
the table provides the run-times (in CPU-seconds) necessary to generate the
architectures.
The results show the benefits of the application-specific architectures. The
increase of the architecture sizes is acceptable. Much more important is the
significant decrease in connection depth, which is a crucial criterion for many time3 The module order with the lowest connection depth is used for a fair comparison.
7 Designing Application-Specific Architectures
number of experiments as well as the length of the experiments) has been increased
until the proposed method did not generate an architecture within 60 min.
Using five operations (a reasonable number of basic operations), the method is
capable of determining architectures for up to
• 32 experiments with an average experiment length of 6.28 modules (the resulting
architecture consists of 30 nodes and 170 edges).
• an average experiment length of 13 modules for 6 experiments (the resulting
architecture consists of 41 nodes and 103 edges).
7.5.2 Comparison to the Ring Architecture
The ring constitutes the commonly used architecture so far. However, this architecture frequently leads to cases where the payload droplet has to traverse the
ring several times (resulting in a high connection depth) or has to be re-injected
by the MPU. Application-specific architectures as generated by the proposed
method overcome these problems. In order to demonstrate this, the proposed
method has been used to generate application-specific architectures for different
sets of experiments. The used experiments are generated out of sequencing graphs
from [102]. More precisely, the operations along each path through the graph are
considered as an experiment. For the generated architectures, both the number of
connections and the number of modules are used as optimization criteria.
Table 7.2 summarizes the obtained results. The first five columns provide the
name, the number of operations (|O|), the number of experiments (||), the
maximal number of considered instances for each module (Max. Inst.), and the
average length of the experiments (Avg. |φ|). Afterwards, the respective results
are reported if the experiments have been realized by means of a ring 3 (denoted
by R) or an application-specific architecture (denoted by G). More precisely, the
number of modules (|V |) and connections (|E|) specify the size of the resulting
architecture. The maximal connection depth (Max. Conn. Depth) and the average
connection depth (Avg. Conn. Depth) state how many connections the payload
droplet has to flow through at most/on average in order to conduct any of the
experiments. The maximal payload re-injections (Max. Re-inj.) and the average
payload re-injections (Avg. Re-inj.) state how often the payload droplet needs
to be re-injected by the MPU at most/on average per experiment. Furthermore,
the table provides the run-times (in CPU-seconds) necessary to generate the
architectures.
The results show the benefits of the application-specific architectures. The
increase of the architecture sizes is acceptable. Much more important is the
significant decrease in connection depth, which is a crucial criterion for many time3 The module order with the lowest connection depth is used for a fair comparison.