176
Network-on-Chip
Input buffer utilization tracks how many buffers in the succeeding router
of the link are occupied. Input buffer age determines how long flits stay in
these input buffers before leaving. These measures reflect resource contention in the succeeding router. The input buffers downstream from the same
link are tracked. Under low network traffic, resource contention is low and
only few buffers are occupied. Flits also do not stay in the input buffers for
long. Hence, both input buffer utilization and input buffer age are low. As
input traffic increases, more flits are relayed between adjacent routers and
resource contention increases, being reflected in higher input buffer utilization and input buffer age. When the network is highly congested, most of the
buffers are filled, and flits are stalled within a router for a long time. Both
buffer utilization and age thus rise dramatically.
Both input buffer utilization and input buffer age track the network congestion point well. They behave like an indicator function that rises sharply
at high network loads. However, compared with link utilization, input buffer
utilization and input buffer age are much less sensitive to changes in traffic. Simulation results show that from lightly loaded traffic to high network
loads, the average buffer utilization only increases by about 0.1. The average
link utilization, however, changes by more than 0.8. Hence, link utilization is
much better at tracking nuances in network traffic.
Link utilization and input buffer utilization are selected as the relevant
measures for guiding the history-based DVS policy, as input buffer age
has similar characteristics as input buffer utilization and is harder to
capture. The link utilization is used as the primary indicator, whereas
the input buffer utilization is used as a litmus test for detecting network
congestion.
6.4.1.1.2 History-Based DVS Policy
Network traffic exhibits two dynamic trends: transient fluctuations and
long-term transitions. History-based DVS policy filters out short-term traffic fluctuations and adapts link frequencies and voltages judiciously to
long-term traffic transitions. It does this by first sampling link (input buffer)
utilization within a predefined history window, and then using exponential
weighted average utilization to combine the current and the past utilization
history:
Weight × Par current + Par past
Par predict =
(6.11)
Weight + 1
where:
Par predicted is the predicted communication link (input buffer) utilization
Par current is the link (input buffer) utilization in the current history period
Par past is the predicted link (input buffer) utilization in the previous history
period
Network-on-Chip
Input buffer utilization tracks how many buffers in the succeeding router
of the link are occupied. Input buffer age determines how long flits stay in
these input buffers before leaving. These measures reflect resource contention in the succeeding router. The input buffers downstream from the same
link are tracked. Under low network traffic, resource contention is low and
only few buffers are occupied. Flits also do not stay in the input buffers for
long. Hence, both input buffer utilization and input buffer age are low. As
input traffic increases, more flits are relayed between adjacent routers and
resource contention increases, being reflected in higher input buffer utilization and input buffer age. When the network is highly congested, most of the
buffers are filled, and flits are stalled within a router for a long time. Both
buffer utilization and age thus rise dramatically.
Both input buffer utilization and input buffer age track the network congestion point well. They behave like an indicator function that rises sharply
at high network loads. However, compared with link utilization, input buffer
utilization and input buffer age are much less sensitive to changes in traffic. Simulation results show that from lightly loaded traffic to high network
loads, the average buffer utilization only increases by about 0.1. The average
link utilization, however, changes by more than 0.8. Hence, link utilization is
much better at tracking nuances in network traffic.
Link utilization and input buffer utilization are selected as the relevant
measures for guiding the history-based DVS policy, as input buffer age
has similar characteristics as input buffer utilization and is harder to
capture. The link utilization is used as the primary indicator, whereas
the input buffer utilization is used as a litmus test for detecting network
congestion.
6.4.1.1.2 History-Based DVS Policy
Network traffic exhibits two dynamic trends: transient fluctuations and
long-term transitions. History-based DVS policy filters out short-term traffic fluctuations and adapts link frequencies and voltages judiciously to
long-term traffic transitions. It does this by first sampling link (input buffer)
utilization within a predefined history window, and then using exponential
weighted average utilization to combine the current and the past utilization
history:
Weight × Par current + Par past
Par predict =
(6.11)
Weight + 1
where:
Par predicted is the predicted communication link (input buffer) utilization
Par current is the link (input buffer) utilization in the current history period
Par past is the predicted link (input buffer) utilization in the previous history
period