blob: 091d4122ecaabd4dcf0c9078d904ca5ed2286088 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
|
DYNAMIC EDP CAPPING IN GENERAL
The goal of dynamic EDP capping is to maximize performance of a system
without violating the peak-current capacity of that system's power
source.
Dynamic EDP Capping makes sense in systems with:
[*] a power source of finite peak-current capacity
[*] one or more controllable variables which have a known
effect on peak current consumption from the power source.
[*] One or more variables whose changes are:
- observable in advance and
- which have a known effect on peak current consumption
from the power source
In a system with only one controllable variable, the control algorithm
is extremely simple. When the observables change, the algorithm solves
for the maximum permissible current associated with the controllable and
then limits the controllable as necessary to keep its current under that
limit.
In a system with more than one controllables, the control algorithm
needs to worry about a strategy which controls the sum of their current
while maximizing performance. There may or may not be a provably correct
algorithm for that. If not, the EDP capping needs to fall back on a
heuristic-based policy for choosing how to spread the pain among the
controllables.
In practice, the selection of controllables and observables is
debatable. The simpler the set, the lower the software &
characterization overhead. However, the simpler the set, the less
accuracy that the observables+controllables provide in estimating the
peak current. The larger the worst-case estimation error, the more
performance must be sacrificed from the controllables in order to avoid
violating the power sources peak-current capacity.
|
