A lot of parameters (optimization criteria) are there, such as processor utilization, power consumption, cache performance, switch cost , etc... but while implementing the scheduling algorithm (particularly genetic algorithm), why do the majority of the researchers use the parameter make-span and flow-time alone ?
Because in industrial settings, the cost of running the algorithm (CPU time, memory, etc.) is small compared to the cost of the industrial plant (for example, is you rent the plant as a unit, the cost is proportional to the make span; if you rent machine by machine, the cost grows as the flow time grows). If you have penalties for delivering late a job and/or rewards for delivering earlier, that time is equal to available time minus flow time Even if you own the plant and machines, you must consider your cost of opportunity (how much money could I earn renting the plat or the machines to others instead of using them )
If you look at the scheduling problems in the context of embeded systems, you can see different problems where the make span is the goal, the objective function but the energy consumption, the size of the memory, most of the parameters you talk about are expressed as constraints.
I am sorry. I thought Dr. Devarajan was asking about job scheduling. I think if you search for "scheduling" (alone or + "cloud") in google scholar, more answers wil be related to job scheduling (industrial settings) than about "scheduling algorithms for operating systems". Stiil, i am not sure about what kind of scheduling is asked for, because the topics section of the question includes cloud computing and parallel computing. If the question is about scheduling many task in many processors, the problem can be considered a traditional job scheduling problem (jiobs= processing tasks, cpus= machines), and my remarks are valid.
Dear Glen, i understand in your last response that you think i want to contradict you. I am sorry about that, but be sure that was not my goal. I just want to give some other contexts or examples where the makespan is NOT the goal or the objective of the scheduling problems. I am agree, that in many contexts, (in grid, cloud), minimizing the makespan is the main objective. But in some other contexts you can also consider the makespan as a constraint : "I want the results of my comptation in 3hours", and optimize on some other criterion : mnimizing the number of CPU used, minimizing the memory consumption, minimizing the energy consumption... Be sure I am agree with your remarks. Dr. Devarajan asks about problems where the makespan was not the primary goal, i just want to give others examples.
If I may add, scheduling in fact has a lineage with operation research. Therefore objectives of scheduling in single or distributed systems are not too different than that of a plant. However, modern computing systems provides oppertunities of various other things to optimize. As Dr LeCun has rightly pointed out there are other things that one might want to optimize depending on the system on which the scheduling problem is applicable. Recently we are looking into scheduling under IoT framework and we bumped into makespan minimization problem.
In solving production scheduling problem, minimization of makespan and completion times are common and popular measure of performances. The reason is- these measures are straight forward way to measure the productivity as they are directly related to production through output or the quantity or amount of raw material processed within a given time, the work done by a production floor in a given period of time, and all other relevant costs or factors like, setup costs, processor utilization, power consumption, cache performance, switch cost are assumed to inherited in these performance measures. However, to measure the overall performance of production, it is obvious that other factors should be considered while optimizing the performance of the system. No doubt, this consideration makes the problem practical, but more complex to solve.
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators