All tested benches have 12 masters and 12 slaves. Com
munication traffic is randomly generated based on several pa
rameters. The frequency of bus access requests is controlled
by the inter-communication interval. For a master, the inter
communication interval is defined as the number of bus cycles
after which a new bus transaction is generated since the pre
vious bus access is granted. The inter-communication interval
is randomly generated following Poisson distribution with av
erage interval values of 1 through 11. The probability that a
slave is accessed is assigned based on the distribution of com
munication distance. Here, the communication distance of a
bus transaction refers to the number of modules between the
master and the destination slave for this transaction. Three
communication distance distributions are used: uniform, Poisson, and exponential. In the uniform distribution, a master has
equal probabilities of accessing all slaves. In the other two distributions, average communication distance is also used as a
parameter.
The experimental results first show that our arbiter implementation can always correctly detect compatible transactions
and establish propagation paths for them. Two metrics are used
to evaluate the bus performance: the effective bandwidth and
the communication latency. The effective bandwidth is defined as the number of finished bus transactions over the total
number of bus cycles used. And the communication latency
refers to the number of bus cycles a master module spent to
obtain the bus access grant. Fig. 5 and Fig. 6 show the effective bandwidth and average latency, respectively, for split
buses with different numbers of segments and different intercommunication intervals. Note that a traditional single access
bus can be viewed as a split bus with only one bus segment.
From the experimental results, we can see that splitting a
bus into multiple segments can always lead to improvement in
both the effective bandwidth and the average latency improvement over traditional single access buses (shown as split buses
with one segment in the figures). Test cases with the exponential communication distance distribution have the highest improvement in terms of both bandwidth and latency. The bandwidth improvement can be as high as 2.3 times and the average
communication latency can be reduced by up to 5 times. The
reason is that there are more bus transactions among nearby
modules in the exponential communication distance distribution, leading to higher possibilities of having compatible bus
transactions. The implication is that the performance of the
split bus architecture can be improved with proper ordering of
modules on the bus. Another observation is that the bandwidth
improvement is dependent on the average inter-communication
interval. This is caused by the fact that there are fewer simultaneous bus access requests when inter-communication interval is large. Therefore, the multiple bus access ability of the
split bus architecture is not fully exploited. However, even in