|
t
4
<
t
<
t
5
), we define level
w
5
to depict the completion time of
the SentiStrength job with
w
6
denoting the shutdown of the VMs
(
t
5
<
t
<
t
6
). In our experimentation,
w
1
and
w
6
were observed
to be very close, indicating almost similar power consumption
in addition to the time intervals for the Hadoop startup
w
2
and
shutdown
w
5
.
3.4. Evaluation and analysis
In this section, we present and analyze results from two per-
spectives, (i) task completion time vs power consumption and (ii)
Effect of scheduling VMs and load balancing.
3.4.1. Task completion time vs power consumption
We study the effect of multitenancy of VMs per PM by analyzing
the tasks completion time versus the power consumption for Sen-
tistrength with various workloads in a Hadoop cluster. Datasets 1
through 5 are used as workloads for Sentistrength application and
is executed on the cluster in various configurations as described in
Table 1
. We compute the ratio of power usage per task in terms
of kilowatts per hour for the one-, two-, four-, and eight-VM con-
figurations.
Table 2
shows the minimum, maximum, and average
power consumption (watts) for the RIoTU testbed. It further shows
the runtimes of various workloads and energy consumption in
kilowatt hours per task for various VM configurations.
The results present a tradeoff between power consumption and
time efficiency for various VM configurations. From this experi-
mentation, we observe that for smaller workloads (datasets 1 and
2), it is more energy efficient to execute the tasks with a single
VM per PM; however, this defeats the purpose of a multitenant
cluster of VMs in terms of efficient utilization of cluster resources.
For larger workloads (datasets 4 and 5), the cluster efficiency in
terms of task completion time improves as we deploy more VMs
to complete the tasks. In all cases, increasing the number of VMs
per server, generally, provides better efficiency in terms of task
completion time when compared to a single VM or physical system
configurations.
On the other hand, with larger workloads, although the job
completion time improves with the increased number of VMs, the
energy cost per job proportionally increases. This can be observed
for Datasets 4 and 5; the task execution time for workloads in
Dataset 4 in the eight-VM configuration is 19% faster at 6210 s
compared to 7420 s in the four-VM configuration; however, the
power consumption is also 28% increased with 0.43 kWh compared
to 0.33 kWh. For Dataset 5 with a 60 GB workload, in a single-
VM configuration, the task execution times are significantly larger
compared to the four-VM and eight-VM configurations. The task
execution time in the eight-VM configuration is 12% faster at
11,240 s compared to 12,580 s in the four-VM configuration. How-
ever, the power consumption is also 32% larger with 0.792 kWh
compared to 0.598 kWh.
Fig. 3
shows the relationship between
the power consumed and the completion rate in terms of time for
datasets 1 to 5.
Increasing the number of VMs per server to 12, we observed
the task completion time for Datasets 4 and 5 is similar to 8-VM
configuration with no benefit in terms of earlier job completion
rate; however, the power consumption increased by 31% to 1.028
kWh. This clearly shows that increasing multitenancy of VMs per
PM does not necessarily improve the task completion rate and may
have a negative effect on overall power consumption and energy
efficiency of the cluster.
3.4.2. Effect of load balancing on energy efficiency
To observe the effect of task scheduling and load balancing on
power consumption of the cluster, we configure Hadoop cluster
to initiate only one data node per VM for a task. The Scheduling
algorithm creates splits in the dataset file and creates map and
reduce tasks. For SentiStrength, Hadoop creates various map tasks
for processing the 64 MB split, the results are written to a file by the
reduce task in parallel. We observe that Hadoop’s task scheduling
and deployment process is CPU and IO intensive. This can be visu-
alized in
Fig. 4
(a) where a spike in the power consumption (power
level
w
3
) during the initial task placement phase is observed. As
the map and reduce jobs initiate, the power consumption of the
cluster normalizes at power level
w
4
. Depending on the workload
on the cluster, the power consumption stays consistent with a
variability of 4% to 5%. In
Fig. 4
(a), we present the progress of map
and reduce tasks for dataset 1 in a 4 VM configuration against the
power consumption in the cluster. Due to the small size of the
dataset, only 4 map tasks and 1 reduce tasks are created conse-
quently launching datanodes on 5 VMs. We observed that 3 VMs
were placed on a PM whereas the rest of the VMs were unevenly
distributed across the cluster. Overall, only 5 VMs were used for
processing the results whereas 10 VMs were idling, negatively
affecting the clusters energy efficiency.
With a larger dataset 4, 192 splits were created subsequently
launching map and reduce tasks on data nodes. In this configura-
tion, since only a maximum of 15 datanodes can be launched in
parallel, the Hadoop scheduler manages the workflow assigning
datanodes to VM. We also observe that the load balancing over the
cluster is not optimal with some VMs idling during execution. This
is evident by the larger variability in power consumption towards
the end of the execution of the workflow as can be seen in
Fig. 4
(b).
3.4.3. Summary
Considering the observations, we deduce that the task place-
ment in Hadoop is not optimal for power efficiency. In fact, the fo-
cus is on early job completion by utilizing the maximum resources
negatively affects the power efficiency of a cluster. It is important
to analyze the power consumption requirements of an applica-
tion in order to improve the power efficiency of the data center.
Task completion times, power-aware scheduling of tasks, sizes of
workloads, multi-tenancy of VMs per server, all play an important
role in optimal utilization of resources. A power-aware workflow
scheduling and task placement strategy is needed to optimize
the resource utilization without compromising the efficiency of
a data center. In the next section, we present a power-aware
workflow scheduling framework that considers optimal resource
utilization in PMs. This optimization is realized by employing APs
that consider the energy-efficiency parameters of tasks for PMs in
the cluster.
Do'stlaringiz bilan baham: |
