Lesandro Ponciano

On the impact of energy-saving strategies in opportunistic grids

Opportunistic grids are distributed computing infrastructures that harvest the idle computing cycles of computing resources geographically distributed. In these grids, the demand for resources is typically bursty. During bursts of resource demand, many grid resources are required, but on other times they remain idle for long periods. If the resources are kept powered on even when they are neither processing their owners workload nor grid jobs, their exploitation is not efficient in terms of energy consumption. One way to reduce the energy consumed in these idleness periods is to place the computers that form the grid in a “sleeping” mode which consumes less energy. We evaluated two sleeping strategies, denoted: standby and hibernate. Resources that comprise an opportunistic grid are normally very heterogeneous, and differ enormously on their processing power and energy consumption. It opens the possibility of implementing scheduling strategies that take energy-efficiency into account. We consider scheduling in two different levels. Firstly, how to choose which machine should be woken up, if several options are available. Secondly, how to decide which tasks to schedule to the available machines. In summary, our results presented a significant reduction in energy consumption, surpassing 80% in a scenario when the amount of resources in the grid was high. Moreover, this comes with limited impact on the response time of the applications.

Assessing Green Strategies in Peer-to-Peer Opportunistic Grids

Opportunistic peer-to-peer (P2P) Grids are distributed computing infrastructures that harvest the idle computing cycles of computing resources geographically distributed. In these Grids, the demand for resources is typically bursty. During bursts of resource demand, many Grid resources are required, but on other occasions they may remain idle for long periods of time. If the resources are kept powered on even when they are neither processing their owners’ workload nor Grid jobs, their exploitation is not efficient in terms of energy consumption. One way to reduce the energy consumed in these idleness periods is to place the computers that form the Grid in a “sleeping” state which consumes less energy. In Grid computing, this strategy introduces a tradeoff between the benefit of energy saving and the associated costs in terms of increasing the job response time, also known as makespan, and reducing the hard disks’ lifetime. To mitigate these costs, it is usually introduced a timeout policy together with the sleeping state, which tries to avoid useless state transitions. In this work, we use simulations to analyze the potential of using sleeping states to save energy in each site of a P2P Grid. Our results show that sleeping states can save energy with low associated impact on jobs’ makespan and hard disks’ lifetime. Furthermore, the best sleeping strategy to be used depends on the characteristics of each individual site, thus, each site should be configured to use the sleeping strategy that best fits its characteristics. Finally, differently from other kinds of Grid infrastructures, P2P Grids can place a machine in sleeping mode as soon as it becomes idle, i.e. it is not necessary to use an aggressive timeout policy. This allows increases on the Grid’s energy saving without impacting significantly the jobs’ makespan and the disks’ lifetime.

Considering human aspects on strategies for designing and managing distributed human computation

A human computation system can be viewed as a distributed system in which the processors are humans, called workers. Such systems harness the cognitive power of a group of workers connected to the Internet to execute relatively simple tasks, whose solutions, once grouped, solve a problem that systems equipped with only machines could not solve satisfactorily. Examples of such systems are Amazon Mechanical Turk and the Zooniverse platform. A human computation application comprises a group of tasks, each of them can be performed by one worker. Tasks might have dependencies among each other. In this study, we propose a theoretical framework to analyze such type of application from a distributed systems point of view. Our framework is established on three dimensions that represent different perspectives in which human computation applications can be approached: quality-of-service requirements, design and management strategies, and human aspects. By using this framework, we review human computation in the perspective of programmers seeking to improve the design of human computation applications and managers seeking to increase the effectiveness of human computation infrastructures in running such applications. In doing so, besides integrating and organizing what has been done in this direction, we also put into perspective the fact that the human aspects of the workers in such systems introduce new challenges in terms of, for example, task assignment, dependency management, and fault prevention and tolerance. We discuss how they are related to distributed systems and other areas of knowledge.

Energy Efficient Computing through Productivity-Aware Frequency Scaling

This paper proposes a new policy for dynamic frequency scaling: productivity-aware frequency scaling (PAFS). PAFS aims at optimizing energy consumptions while still satisfying performance requirements of a given application. In contrast to the commonly-used on demand frequency scaling, PAFS may keep the processor in a power save state even in high CPU-usage situations. This will be the case as long as the application (or set of applications) for which productivity is to be preserved presents acceptable performance (e.g., as stablished by a QoS contract). Our experiments show savings of up to 23.65% in energy consumption when compared to the commonly used on demand DFS policy with no performance degradation for the productivity metric. PAFS is, therefore, binded to a single or a set of applications running in a machine. Nevertheless, compared to previous approaches to application-specific frequency scaling, PAFS does not require modifying the application or a calibration process. PAFS requires only a productivity metric which may already be exported by an application (e.g., through a log file, such as response time or throughput in an Apache web server) or which may be computed through a simple program or script.

Adaptive Task Replication Strategy for Human Computation

Human computation systems are distributed systems in which the processors are human beings, called workers. In such systems, task replication has been used as a way to obtain results redundancy and quality. The level of replication is usually defined before the tasks start executing. This approach, however, generates the problem of defining the suitable task replication level. If the level of replication is overestimated, it is used an excessive amount of workers and, therefore, there is an increase in the cost of executing all tasks. On the other hand, if the level of replication is underestimated, a desired level of quality cannot be achieved. This work proposes an adaptive replication strategy that defines the level of replication for each task during execution time. The strategy is based on estimations of the degree of difficulty of tasks and the degree of credibility of workers. Results from simulations using data from two real human computation applications show that, compared to non-adaptive task replication, the proposed strategy reduces the number of replicas substantially, without compromising the accuracy of the obtained answers.

BitTorrent traffic from a caching perspective

BitTorrent currently contributes a significant amount of inter-ISP traffic. This has motivated research and development to explore caching and locality-aware neighbor selection mechanisms for costly traffic reduction. Recent researches have analyzed the possible effects of caching BitTorrent traffic and have provided preliminary results on its cacheability. However, little is known about the specifics of caching design that affect cache effectiveness and operation, such as replacement policy and cache size. This study addresses this gap with a comprehensive analysis of BitTorrent caching based on traces of user behavior in four popular BitTorrent sites. Our trace-driven simulation results show differences in BitTorrent traffic caching compared to that of the Web and other peer-to-peer applications. Differently from Web and other peer-to-peer caching, larger caches are necessary to achieve similar caching effectiveness in BitTorrent traffic. Furthermore, in BitTorrent caching, the LRU replacement policy that takes the temporal locality into account shows the best performance. We also use a locality-aware neighbor selection mechanism as a baseline to evaluate the LRU caching effectiveness. We find that LRU caching can provide greater traffic reduction than locality-aware neighbor selection in several scenarios of cache size and number of ISP clients.