Using Voltage and Frequency Adjustments toManage Dynamic Power - Computer Architecture Techniques for Power-Efficiency - page 38

throughput of the corresponding domain. 3 When a queue fills up—its utilization increases—the
throughput of its domain is low. The analysis of Semeraro et al. revealed that (i) decentralized
control of the different domains is possible and (ii) the utilization of the input queues is a good
indicator for the appropriate frequency of operation.
Based on the observations of their analysis, Semeraro et al. devised an online DVFS con-
trol algorithm for multiple domains called Attack/Decay. This is a decentralized, interval-based
algorithm. Decisions are made independently for each domain at regular sampling intervals.
The algorithm tries to react to changes in the utilization of the issue (input) queue of each do-
main. During sudden changes, the algorithm sets the frequency aggressively to try to match the
utilization change. This is the Attack mode. If the utilization increased by a significant amount
since the last interval, the frequency is also increased by a significant factor. Conversely, when
utilization suddenly drops, frequency is also decreased. In the absence of any significant change
in the issue queue, frequency is slowly decreased by a small factor. This is the Decay mode.
The algorithm tries to balance the utilization of the various domains independently by
varying their speeds. This, however, does not account for a natural change in the performance
of the program. To capture this effect, the performance of the processor in terms of IPC is
tracked from interval to interval. If there is an inherent change in the IPC that is not related
to frequency adjustments in the domains, then no frequency changes are allowed for the next
interval. The IPC is the only global information needed in this algorithm.
On average, across a wide range of MediaBench, Olden, and Spec2000 benchmarks, their
algorithm achieved a 19% reduction (from a non-DVFS baseline) in energy-per-instruction
and a 16.7% improvement in energy-delay product. The approach incurred a modest 3.2%
increase in cycles per instruction (CPI). Interestingly, their online control-theoretic approach
was able to achieve a full 85.5% of the EDP improvement offered by the prior offline scheduling
approach. Wu et al. extended the online approach using formal control theory and a dynamic
stochastic model based on input-queue occupancy for the MCDs [ 228 ].
3.4.2 Dynamic Work-Steering for MCD Processors
As an alternative to dynamic voltage/frequency scaling of multiple clock domains, one can
statically provide multiple components for the same function clocked at different frequencies.
For example, one can provide a fast and power-hungry pipeline and a slow but power-efficient
pipeline. With two such pipelines, the problem is no longer about selecting domain frequencies
to eliminate stack, but instead is about steering instructions to the appropriate slow or fast
3 The occupancy of the instruction queue is also used for resizing it to reduce its switching activity factor as we discuss
in Section 4.6.3.