Moldable Job Allocation for Handling Resource Fragmentation in Computational Grid - Cloud, Grid and High Performance Computing: Emerging Applications - page 67

Figure 2. Effects of inexact runtime estimation under uniform distribution

tive scaling up and down policy delivers the best
performance when the threshold value is 2. Figure
1 shows that moldable job allocation in general
can improve the overall system performance
several times, compared to the traditional alloca-
tion policy sticking to a job's original amount of
processor requirement. Moreover, the improved
moldable job allocation policies presented in
this paper can further improve the performance
significantly with the aid of runtime estimation.
For the original moldable job allocation policies,
allowing scaling up parallelism cannot improve
system performance further in addition to scaling
down parallelism in terms of average turnaround
time. However, for the improved moldable alloca-
tion policies, scaling up parallelism does improve
the system performance delivered by the policy
which scales down the parallelism only. Overall
speaking, the conservative scaling up and down
policy with runtime estimation outperforms the
other policies.
The studies in Figure 1 assume that users al-
ways provide exact estimations of job execution
times. However, this is by no means possible in
real cases. Therefore, we performed additional
simulation studies to evaluate the stability of

the moldable job allocation policies when users
provide only inexact estimations. The results are
presented in Figure 2. The error range of estima-
tion is relative to a job's actual execution time.
Figure 2 shows that sometimes small estimation
error might even lead to better performance than
exact estimation such as the case of conservative
scaling up and down with a 20% error range. In
general, a larger error range results in degraded
performance. However, up to 90% error range, the
improved moldable job allocation policies with
runtime estimation still outperform the original
moldable allocation policies, compared to Figure
1. The results illustrate that the proposed moldable
job allocation policies are stable and practical.
The simulations for Figure 2 assume the esti-
mation errors conform to the uniform distribution.
Figure 3 presents another series of simulations
which evaluate the cases where the estimation
errors conform to the normal distribution. The
results again show that sometimes larger error
ranges lead to better performances. Moreover,
Figure 3 indicates that the moldable job allocation
policies perform even more stably under the
normal distribution of estimation errors, compared
to Figure 2.