Distributed resource management and scheduling, including energy-aware techniques By: Greg Boyarko, Jordan Sutton, and Shaun Parkison CS 455: Introduction to Distributed Systems Computer Science Department, Colorado State University.

Why is this problem important? Concerned with making sure that a user/client can access remote resources as easy as it can access the local resources. Concerned with being energy efficient. Advantages: System scalable DSM is usually cheaper than using multiprocessor system provides large virtual memory space Describe where this research is being used, how this plays a role in our daily lives, etc

Problem characterization Today’s data-centers are consuming nearly 2% of the global energy In a report by the EPA, it is estimated that the power consumed by servers and data centers has more than doubled between the years 2000 and 2006 Until recently, high performance has been the sole concern in data center deployments, and this demand has been fulfilled without paying much attention to energy consumption Resource management and scheduling needs to be adaptive to changes in resources and users requirements, while remaining scalable, controllable, and measurable A decrease in energy typically means a lower clock-cycle, or increased latency This is a technical description of the problem. Your audience is your peers so express it in a way that they can appreciate.

Trade-off space for solutions in this area Most techniques to implement a more energy efficient means of job scheduling usually comes at the cost of decreased performance This can lead to a direct impact on a company's revenue One of the major sources of energy consumption in a data center is the actual cooling of the machines Mapping tasks to machines in heterogeneous data center systems has been shown to be, in general, an NP-complete problem Often the problem space is big enough that optimal solutions might be computationally intractable. In some cases, the memory requirements may be too high. Distributed solutions may entail far too many message exchanges. Add in consumer demands such as fast response times and things get even trickier. To cope with such scenarios, solutions often go for the “good enough” approach. The trade-off space describes which element of the solution space was traded off for achieving a certain objective. In cloud settings for example, solutions often trade-off consistency for availability; if there are failures in the system, you will still be able to access the service but consistency might be off – it is always a good idea to check if the quantities are correct in your online shopping cart before you complete the checkout.

Dominant approaches to the problem Energy aware techniques like Horizontal scaling, Vertical scaling, and Dynamic Voltage/Frequency Scaling (DVFS) Scheduling techniques that best take advantage of a processors computing power while reducing idle time like Best Fit Hole (BFH), NOUR Best Fit Hole (NOUR), Earliest Finish Time First (EFTH), and Optimistic Load Balancing (OLB) This is where you quickly summarize your survey while giving a feel for the key features in each approach that you surveyed.

Dominant approaches to the problem CoolEmAll, a project that aims to reduce energy costs and maximize energy efficiency by optimizing cooling and workload management Utilization of cellular automation to map different tasks to cells that are governed by rules generated from a genetic algorithm Second slide for dominant approaches...

Insights Gleaned DVFS was able to minimize the number of computing servers and time Genetic Algorithms, which mimic natural selection thus finding an optimal solution the same way nature does, were the most effective at energy-conservation, though they had the highest overhead for scheduling These are the key ideas that underpin the best approaches that you surveyed. These are things that you did not know before you started the survey. The best solutions are the ones that you may have not thought of, but seem incredibly obvious once you have seen them.

What the problem space in the future would look like Energy-aware scheduling algorithms Include energy use as a goal instead of just throughput or speed Sometimes these goals will coincide nicely, due to all trying for efficiency If they don’t coincide, you must choose what is more important Volunteer computing Idle computers can be used to help with other jobs Jobs can be kicked out if computer is needed by main user Schedule determines when computers are not likely to be needed Machine learning can be used to develop optimal schedule This is a thought experiment. You will be looking ahead and visualizing a future where there could be proliferation of certain types of devices, new types of services, changes in usage patterns, etc. This slide will describe how you expect the problem space to evolve in the future.

What the problem space in the future would look like Price-aware scheduling algorithms Large consumers of electricity may purchase at real-time prices Scheduling can determine when best time to buy is Growth of renewable resources will make prices fluctuate wildly Good schedule is essential, as peak prices can be far above average

Trade-off space and solutions in the future Scheduling optimization can have trade-offs What factors do you consider most important? Geographical Location Proximity vs Heat Environmental impact Smart Devices Privacy concerns This where you will be proposing “rough” solutions based on the problem space in the future. You will be applying what you have learned so far.

Conclusion Energy efficient means of job scheduling come at the cost of decreased performance or increased latency. Half of the actual energy costs come from cooling machines. Optimizing idle time could help to dramatically reduce energy usage. Essential to find a more energy efficient means of data transfer, with minimal impact on performance.