Embedded Software for SoC - Grupo de Mecatrônica EESC/USP

Adaptive Checkpointing with Dynamic Voltage Scaling 459
Table 33-6. (a) Variation of energy consumption with for D = 10000, c = 10, and k = 10.
U
Fault arrival rate
Energy consumption
Poisson-arrival
k-fault-tolerant
ADT_DVS
0.60
2.0
4.0
6.0
8.0
25067
25574
25915
26277
26327
26477
26635
26806
21568
21642
21714
22611
Table 33-6. (b) Variation of energy consumption with U for D = 10000, c = 10, and k = 10.
U
Energy consumption
Poisson-arrival
k-fault-tolerant
ADT_DVS
5.0
0.10
0.20
0.30
0.40
0.50
4295
8567
12862
17138
21474
4909
9335
13862
17990
22300
2508
4791
7026
9223
15333
voltage and the number of computation cycles over all the segments of the
task.) This is expected, since ADT_DVS attempts to meet the task deadline
as the first priority and if either or U is high, ADT_DVS seldom scales
down the processor speed.
5. ADAPTIVE CHECKPOINTING FOR MULTIPLE TASKS
We are given a set of n periodic tasks, where task is
modeled by a tuple is the period of is the relative
deadline and is the computation time under fault-free conditions.
We are also aiming to tolerate up to k faults for each task instance.
The task set is first scheduled off-line with a general scheduling algorithm
scheme under fault-free conditions. Here we employ the earliest-deadline-first
(EDF) algorithm [11]. A sequence of m jobs
is obtained
for each hyperperiod. We further denote each job
as a tuple
where is the starting time, is the execution time and is
the deadline for All these parameters are known a priori. Note that is
the absolute time when starts execution instead of the relative release time,
execution time is equal to that for the corresponding task, and job deadline
is equal to the task deadline plus the product of task period and corresponding
number of periods. In addition, since we are employing EDF, we have
Based on the job set
we develop a checkpointing scheme that inserts