The info policy specifies the load information made avilable to the job placement decision makers. This info consists of :

• Load information of all or a subset of PE's
• Load information is distributed periodically or on-demand at oad balancing time
• The load info is collected at a central PE or distributed among the participating job placement decision makers.

 Transfer policy
  The transfer policy determines the suitablity of a process for task relocation. The size of the task can be quantified as the task execution time. Thus the parameters used are :

• Host load : A process is eligible for transfer only if the host load is greater than a threshold value
• Process size : Check for size greater than threshold value, thus making the advantage of task relocation outweigh the transfer overhead
• Process resource needs : Since a process may not be executable on all possible machines or its execution times may vary, we need to clearly specify this requisite

  Placement policy   This determines the PE to which a process is assigned, where the selection is made either based on Minimum Load PE selection or Low Load PE selection(first possible PE with load less than threshold).

  The issue of Load indices, a measurement of the PE's load can be based on the size of the ready queue of the PE. Exact estimaton of the size of the process is not possible due to decision making only at load balancing time.

Another issue is the Task Migration, where we can suspend the execution of a task on one PE, transfer it to antother PE and resume the execution.

A third issue is that of Thrashing, or instability, that arises when several overloaded PE's transfer their tasks to an underloaded PE causing it to become overloaded, resultingf in it passing the tasks to some other PE, causing thrashing. We can solve this by :

• Selecting the lightly loaded PE's for job transfer in a random way, instead of choosing the PE with the imnimum load.
• Randomized searching for a target PE, for allowing some time ofor propagation of the load info to the job placement decision makers.

Classification of Load Balancing Methods

  There are basically 3 different criteria for classification :

• Sender or receiver initiated
• Static or adaptive
• Central or distributed info and placement decision

 Sender or receiver initiated : An algorithm is said to be sender-initiated if a local host makes a determination as to where a generated task or arriving task is to be executed. The queues of ready jobs tend to form at the target PE's. Job transfer decisions are made at task arrival time. In a receiver initiated process, a server or target PE determines which jobs at different sources, it will process. The ready jobs tend to queue at the source PE's. with job transfer decisions made at task completion time.

 Static or Adaptive: In the former, the transfer and placement policies are based on the information of the average behaviour of the system. These algorithms do not depend on the current state of the system at the time of decision making. Random and round-robin placement policies are examples of static decision making processes. Adaptive schemes react to changes in the system state. The complexity of these schemes is an obvious drawback as they require runtime load information and state collection activities. Threshold scheduling are examples of this .

  Central or Distributed information and placement decisions : There are three categories :

• Centralized policies :A PE designed as a central scheduler, receives updated load information from all other hosts and assembles them into a load vector. For the placement policy, when a host decides to transfer a job, it sends a request to the central scheduler which selects the target host using the load vector and informs the source host of its choice. These policies reduce system overhead due to the load distribution information.
• Distributed policies : Here each host autonomously constructs its own local load vector by collecting the load info from the other PE's. The placement decisions are made locally. These policies speed up the process but generate large communication overhead due to load info distribution.
• Mixed policies : A combination of central/distribution policies can be used. For example, we have aa central load information policy along with distributed placement policy.

Future directions

  The major drawback for dynamic loasd balancing is the runtime overhead for load info distribution, making placement decisisons and transferring a job to a target host. The goal is to achieve less overhead, by using effective load index measures and efficient load distribution methods. Hierarchical system organizations with local load info distribution and local load balancing policies seem to be ideal for more efficiency. An extension can be made in future to message-passing multiprocessor systems.