Information Technology Reference
In-Depth Information
ing high performance. There are three goals of
the load balancing policy shown in the following:
On the other hand, according to the manage-
ment approach, load balance policies could also
be categorized into two types.
•
To distribute the workload from high-load-
ing sites to low-loading sites.
Centralized Policy
•
To maximize the resource utilization.
•
To minimize the job execution time.
The centralized load balancing policy adopts one
computing node to be the resource manager and
makes the load balance decision. The centralized
resource manager manages global resource infor-
mation by collecting information from all sites.
The global resource information facilitates the
resource manager to allocate resources. Therefore,
the centralized policies manage resources easily
and achieve better performance. However, central-
ized resource manager could be the bottleneck of
the system; moreover, it may become the single
point of failure.
According to the decision making approach,
load balancing policies can be categorized into
two types.
Static Policy
The static load balancing policy (Pan et al.,
2007) makes the balance decision by the resource
information before executing jobs. It is easy to
implement the static load balancing policy, and
the overhead of implementing the static policy
is lower than that of implementing the dynamic
policy. However, it is more difficult for the static
policy to obtain the optimal performance due to
that it can not adjust the decision at runtime. In
the high variation system, the performance of the
static policy is very poor.
Distributed Policy
The distributed load balancing policy allows every
computing site in the distributed system to make
load balance decisions. In addition, the computing
site only needs to collect the information from its
linked sites. Although the cost of obtaining and
maintaining the dynamic system information is
very high, the distributed policy still could make
the decision successfully when one or more sites
join or leave the system. Therefore, the stability
of the distributed policy is better than that of the
centralized policy. The distributed policy is usu-
ally used in distributed system. Shah et al. (2007)
propose a decentralized load balancing algorithm
which employs the job arrival rates and the job
response for making load balancing decisions. Lei
et al. (2007) make the load balancing decisions
according to the CPU and memory status. Tang
et al. (2008) improve the system stability through
the resource-constrained load balancing control-
ler. Liang et al. (2008) propose an adaptive load
balancing algorithm which makes the workload
of all nodes as evenly as possible. Subrata et al.
(2008) propose a decentralized game-theoretic
Dynamic Policy
The dynamic load balancing policy (Chen et al.,
2008; Duan et al., 2008) makes decisions of the
resource allocations by the runtime information.
Although the dynamic policy (e.g., JRT (Wu et
al., 2008), Max-Min (Ali et al., 1999), Min-Min
(Ali et al., 1999) and RESERV (Vincze et al.,
2008)) brings better performance, it is difficult
to be implemented due to that it needs to collect
the dynamic information for making the optimal
decision. In general, the dynamic policy has the
better performance than that of the static policy.
In addition, the dynamic policy can maximize
the system performance in the high variation
environment.
Search WWH ::
Custom Search