Information Technology Reference
In-Depth Information
Fig. 2.
GPU Discrete Firefly Algorithm Model
to the greatest possible extent. The local populations
P
and
temp
are stored
in global memory. Due to the size of solutions that manage the use of shared
memory, it becomes less feasible in our model, since the storage is limited when
compared with solution size. However, this approach can be studied in the future
with other kind of combinatorial problems or as an improvement in the model.
In this work, we have considered the number of kernels as the main criterion
to assess how well an algorithm can be parallelized. Due to the complexity of
some operations that are completely different from each other, we have tried to
write simple and small kernels since the kernel launch cost is negligible with the
operations to perform and less registers are used.
On the other hand, the performance of a nature-inspired algorithm largely
depends on the quality of its random number generations. For this work, we have
utilized a Mersenne Twister random generator approach [35]. We have employed
a global seed pass at the beginning of the GPU-DFA execution; then, each thread
is initialized with different seed values (by modifying the initial global seed) in
the device. Finally, they are invoked sequentially by each thread for subsequent
random number generations.
4 Experimental Settings
In this section, two well-known permutation combinatorial problems are intro-
duced in order to test the GPU-DFA.
