Biomedical Engineering Reference
In-Depth Information
Figure 12.1
Parallel implementation of the Granger analysis--communication between input node,
master node and worked nodes.
For output consistency, the master node was also assigned the task of writing
all the results to the output files. The worker nodes send the results to the master
node before requesting more work. The master node collects these results and
writes to the output files.
For handling the input latency, we used the standard pipelining technique to
overlap the input and preprocessing with computation. In a typical analysis, there
are multiple sessions of data to be processed. Once the input node has finished
reading, preprocessing, and broadcasting the data, it does not wait for the worker
nodes to finish their computations. Instead, it moves to reading the next session,
while the worker nodes carry out the computations. This sequence of events is
depicted in Figure 12.2. Once the next session is read and preprocessed, the input
node waits for the worker nodes to finish the computations and gets the prepro-
cessed data for the next session. The compute time is typically larger than the time
Figure 12.2
Timing relationship between input node and worked nodes.
Search WWH ::
Custom Search