Load balancing is the problem of assigning tasks to a plurality of resources in a way such that the assignment is optimal in some sense. This problem has been of significant industrial importance where jobs, that require same or different amount of time on various machines, have to be assigned to machines, to balance the load, such that the maximum amount of time taken to complete all the tasks is minimized. An additional condition of this problem is that every job, mentioned above, is assigned to only one machine for its completion.

The method of optimization of the load balancing problem described above is carried out in two phases. In the first phase some tasks are assigned to resources, they are best on, on a per resource basis and in the second phase resources are chosen for the remaining tasks such that the length of the schedule is minimized. For the explanation of the heuristic algorithm let us consider a simple example. Table 1 shows a table of completion times of tasks on their respective resources. Each row of the table corresponds to a task and each column of the table corresponds to a resource. The value of each cell in the table corresponds to time required by a task for its completion on its respective resource. â€œInfâ€ means that the resource is unavailable for its corresponding task.

TABLE 11020Inf301020Inf1010502030204060302020101020403010102030InfInf20

The description of the two phases of the heuristic load balancing algorithm is given below:

Phase 1

In this phase first we find the minimum of each row, sum up all the minimums and divide the sum by the number of resources. For the example, the minimum values of the rows are highlighted below in Table 2.

TABLE 2

The sum of the highlighted minimum values is 150 and dividing that by the number of resources (equal to 3) we get 50. The maximum of average of sum of minimum completion times and the maximum of the minimum completion times is taken as the lower bound of schedule. In this case the average of the sum of minimum completion times computed above is 50. The maximum of the minimum completion times highlighted in the table above is 30. Hence the lower bound of the schedule is MAX (50, 30)=50.

Next we highlight the minimum completion times of every task and sum up the minimum completion times for all the resources. The minimum completion time values of every task are highlighted in Table 3 below.

TABLE 3

The sum of the minimum completion time values for all the resources is {50, 70, 60}. Now we sort the columns of the table in increasing order of completion time sums computed above. If two columns have the same completion time sum then the column having more variation of values comes before the column having less variation. In the example, based on the sorting criterion mentioned above, we swap columns 2 and 3 to make the completion time sums as {50, 60, 70} as shown in Table 4.

TABLE 4

The two values in brackets besides the completion time value in each cell are the (TaskID, ResourceID) of the original table, Table 1.

Having done all this we create a separate table of all the minimum values where minimum values of each column are sorted in increasing order of their completion time values. In a column if two cells have the same value then the cell of the row having less number of infinity values comes above a cell of a row having more number of infinity values and if the number of infinity values is also the same then the cell of the row having minimum completion time sum comes above the cell of the row having larger completion time sum. Table 5 with sorted minimum values is shown below.

TABLE 5

Having obtained Table 5, we proceed from column 1 to 3. We pop off values from the end of the column we are considering and assign popped off tasks to their respective resources. We should only pop off values from the ends till the sum of popped off values of a column is less than or equal to the lower bound we computed before (value 50 for the example in this document). When we pop off values from the bottom of Table 5 and schedule the task on the respective resource then we should mark that task as assigned. In the case of this example the assignment of tasks to resources is in the following order:

(5, 1), (1, 1), (9, 1), (7, 1), (10, 3), (6, 3), (3, 3), (4, 2), (2, 2)

Notice above that after task 3 is assigned to resource 3 the sum of popped off values (20+20+10) in that column equals to the lower bound hence we do not pop off the next value in column 2. Instead we start popping off values from column 3. Also note that after scheduling task 4 to resource 2 we pop off the next cell value without scheduling task 6 was earlier scheduled to resource 3.

At the end of previous step all the tasks except task 8 is assigned to a resource. The original Table 1 looks as below where tasks assigned to their respective resources are highlighted.

TABLE 7

This completes phase 1 of the load balancing algorithm. Next phase is scheduling based on minimization of the length of the schedule.

Phase 2

After phase 1 we compute the total length of the schedule on each resource by summing up the cell values of assigned tasks in the respective columns. For the example the length of the schedule on the resources is {50, 30, 50}. Also, we remove all those rows whose corresponding tasks have been scheduled on some resource. For the example in this document after removing assigned rows we are just left with task 8 as shown in Table 7.

TABLE 6

We sort the rows of the table of unassigned tasks according to decreasing completion time sum. A row with more â€œInfâ€ is considered to have a higher sum than a row with less number of â€œInfâ€ . If two rows have the same number of â€œInfâ€ 0 then the row with a greater sum of values comes above a row with lesser number of values. Once the sorting is done we go from top to bottom of the table of unassigned tasks. We pick a resource for the current task such that the length of the schedule on a resource is minimized. If more than one completion time minimizes the total length then the smallest completion time minimizing the total length is chosen. In the case of the example both completion times 30 and 10 of Table 7 minimize the length of schedule with the minimum length in both cases being 60 (30+30 or 50+10). However because 10 is less than 30 therefore task 8 is assigned to resource 3.

The final solution to the scheduling problem in this example by the load balancing algorithm is shown below in Table 8.

TABLE 8