DARPA ULTRALOG Final Report - Industrial and Manufacturing ...
DARPA ULTRALOG Final Report - Industrial and Manufacturing ...
DARPA ULTRALOG Final Report - Industrial and Manufacturing ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
5<br />
LI i<br />
w i( t ) = wi<br />
= ω n( i ) for all t ≥ 0 , (9)<br />
LI<br />
∑<br />
p∈K<br />
n(<br />
i )<br />
p<br />
T<br />
LB<br />
s<br />
ωn<br />
+ ωn<br />
= Max<br />
n∈N<br />
ω<br />
n<br />
s<br />
∑<br />
i∈<br />
K n<br />
LI<br />
i<br />
(12)<br />
where n(i) denotes a node in which component i resides.<br />
Theorem 2. T s´ equals to T s LB under proportional allocation.<br />
Proof. RA i (t) is more than or equal to assigned weight proportion<br />
as:<br />
w i ( t )<br />
RA ( t ) ≥ for t ≥ 0 . (10)<br />
ω<br />
i<br />
n( i )<br />
Proof. RA i (t) becomes:<br />
RA i( t ) ≥<br />
ω<br />
w ( t )<br />
n( i )<br />
i<br />
+ ω<br />
s<br />
n( i )<br />
for t ≥ 0 . (13)<br />
Suppose a component i receives its tasks at a constant interval<br />
of T LB /L i . Then, under proportional allocation, S i (t) is less<br />
than or equal to T LB /L i over time as shown in (11).<br />
P<br />
i<br />
=<br />
∫<br />
wi<br />
=<br />
ω<br />
n( i )<br />
LB<br />
T<br />
⇒<br />
L<br />
i<br />
t+<br />
Si( t )<br />
t<br />
S ( t ) =<br />
i<br />
RA ( τ )dτ<br />
≥<br />
≥ S ( t )<br />
i<br />
i<br />
LI<br />
∑<br />
i<br />
LI<br />
p∈K<br />
n( i )<br />
p<br />
∫<br />
for t ≥ 0<br />
t+<br />
Si(<br />
t<br />
t )<br />
LI<br />
Si( t ) ≥<br />
T<br />
w i( t )<br />
dτ<br />
ω<br />
n( i )<br />
i<br />
LB<br />
S ( t )<br />
i<br />
(11)<br />
So, any component can complete by T LB <strong>and</strong> generate tasks at<br />
a constant interval of T LB /L i from t=T LB /L i (first task<br />
generation time) under proportional allocation when it<br />
receives tasks at a constant interval of T LB /L i from t=0 (first<br />
task arrival time). As tasks are infinitesimal <strong>and</strong> root tasks<br />
increase task availability, each component can receive<br />
infinitesimal tasks at a constant interval in 0≤t≤T LB or more<br />
preferably, <strong>and</strong> complete at less than or equal to T LB . So, the<br />
network completes at T LB under proportional allocation. <br />
From Theorem 1 we can conjecture that a network can<br />
achieve a performance close to T LB under proportional<br />
allocation in the limit of large number of tasks. We propose the<br />
proportional allocation as an optimal resource allocation<br />
policy. Though the proportional allocation is localized, the<br />
network can maximize the utilization of distributed resources<br />
<strong>and</strong> achieve desirable performance. Coordinated resource<br />
allocation throughout the network emerges as a result of using<br />
the load index as global information. If nodes do not follow the<br />
proportional allocation policy, some components can receive<br />
their tasks less preferably resulting in underutilization <strong>and</strong><br />
consequently increased completion time as have shown in the<br />
previous subsection.<br />
Another important property of the proportional allocation<br />
policy is that it is itself adaptive. Suppose there are some<br />
stressors sharing resources with the components. We denote<br />
ω s n as the amount of shared resource by a stressor in node n.<br />
Then, the lower bound performance T LB s under stress is given<br />
by (12). We denote the completion time under stress as T s´.<br />
Then, (11) results in (14) under proportional allocation.<br />
LB<br />
Ts<br />
L<br />
i<br />
≥ S ( t ) for t ≥ 0<br />
(14)<br />
i<br />
Therefore, the network completes at T LB s under proportional<br />
allocation. <br />
Theorem 2 depicts that the proportional allocation policy is<br />
optimal independent of the stress environments. Though we do<br />
not consider them explicitly, the policy gives lower bound<br />
performance adaptively. This characteristic is especially<br />
important when the system is vulnerable to unpredictable stress<br />
environments. Modern networked systems can be easily<br />
exposed to various adverse events such as accidental failures<br />
<strong>and</strong> malicious attacks, <strong>and</strong> the space of stress environment is<br />
high-dimensional <strong>and</strong> also evolving [26]-[28].<br />
C. Adequacy criterion<br />
The arguments we have made hold in the limit of large<br />
number of tasks. As the term “large” is obscure we need to give<br />
it a concrete definition. We define it with an adequacy criterion,<br />
by which one can evaluate if the desirable properties of the<br />
proportional allocation hold for a given network. For this<br />
purpose we characterize upper bound performance of a<br />
network under proportional allocation.<br />
Theorem 3. Under proportional allocation a network’s upper<br />
bound T UB of completion time T is given by:<br />
T<br />
UB<br />
= T<br />
LB<br />
+ Max Max<br />
e∈E<br />
j∈Se<br />
∑<br />
i∈j<br />
[ P<br />
∑<br />
LI<br />
i<br />
p∈K<br />
n(<br />
i)<br />
p<br />
/ LI<br />
i<br />
] , (15)<br />
where E denotes a set of components which have no successor<br />
<strong>and</strong> S e a set of task paths to component e. A task path to<br />
component e is a set of components in a path from a<br />
component with no predecessor to component e <strong>and</strong> does not<br />
include component e.<br />
Proof. From (11) we can induce the lowest upper bound S i UB of<br />
S i (t) as: