Adaptive Control of Redundant Task Execution for Dependable Volunteer Computing - Cloud, Grid and High Performance Computing: Emerging Applications - page 147

Table 1. Simulation parameters for LFPD policy and the greedy dispatch policy

Skype Trace Microsoft PCs Trace

Number of Workers
Number of Tasks

2,081
80,000

51,663
2,000,000

Mean Task Process Time
Number of Task Groups
Idle Worker Inquire Interval

1250, 2500, 5000, 10000 seconds
5, 10, 20
200 seconds

for about 28 days. Skype is a peer-to-peer VoIP
software that connects thousands of volatile peers.
The Microsoft PCs trace data set (Bolosky, 2000)
stores the availability data of 51,662 desktop PCs
within the Microsoft corporation network for 35
days. The volatile peers in the peer-to-peer network
and desktop PCs in the corporation network are
two typical worker types for volunteer computing.
In a heterogeneous environment, the perfor-
mance of each worker is different. To simulate such
an environment, worker's performance parameters
are generated with a power-law distribution. As
this work focuses on the computation-intensive
problems that satisfy `` computation time ≫ data
transfer time ,'' the communication cost is not
considered in the simulation.
The simulation parameters are as follows:

• Idle Worker Inquire Interval: an inquire

interval of idle workers that received the
``sleep'' message. The ``sleep'' message is
only used in the greedy dispatch policy.

• Number of Task Groups: the number of

task groups in the computing job. It is the
factor of inter-task dependency.

The simulation parameters are listed in Table 1.
Since the same mean task process times are
used to evaluate the two trace data sets, the dif-
ferent availability characteristics make it hard to
compare the evaluation results of these two trace
data sets. Thus, the Microsoft PCs trace data set
is modified to have the same mean TTF as the
Skype trace data set. The basic statistical proper-
ties of these two trace data sets are shown in
Table 2.

• Number of Workers: the number of work-

ers in the platform. It is the number of
peers in the trace data sets.

Performance Evaluation

• Number of Tasks: the number of tasks for

the computing job.

The two dispatch policies are evaluated with
different parameters and different resource avail-
ability trace data sets. The total process time of
the computing job for different combinations are
compared and discussed. To simply the discussion,
all the results are normalized with the correspond-

• Mean Task Process Time: mean process

time of a task. The process time of a task
on a worker depends on the performance
parameter of his worker.

Table 2. Summary of the basic statistical properties of the data sets

Skype Trace Microsoft PCs Trace

Mean TTF (seconds)
Mean Down Time (seconds)
Average percentage of online node

55,125
51,509
33.15%

55,125
15,906
81.24%