�ݺ�ߣ

Editor's Notes

• #3: Once the federation we are dealing is completely decentralized one of the greatest challenges we face is the promotion of cooperation among the participants. In P2P networks, in a first glance, peers don’t know each other and are assumed to be selfish and to have economic incentive to become free riders, which are the non-reciprocators peers. Therefore, some type of incentive mechanism must exist in order to maximize the benefits of the peers staying in that federation, otherwise, they would leave it. A federation will provide good results to a peer only if he takes efficient decisions, which includes not donating to free riders. Thus, our incentive mechanism should help the peer to answer the following questions that they naturally have: To whom should I donate? (here, the mechanism should point out collaborative partners) How much should I donate? (once I do not trust anybody, should I donate all my exceeding resources to this unknown peer, or should I increase this ammount gradually?)
• #4: Network of Favors (NoF): Ensures the prioritization of favors to colaborators ; In NoF, each peer keeps a local record of his past interactions with other peers. For example: peer D provided 300 units of favor to peer A, and condumed 150 units of favor from peer A, which gives him a credit of 150 units of favor with peer A. A unit of favor could be a measure that involves a processing power (which could be described as a flavour of the instance) multiplied by the time it was lent. As you can note, this table is ordered by the credit, and whenever peer A has exceeding resources, he will always prioritize the peers in the top of this table. This is how the NoF reward colaborative peers.
• #5: Whenever there is contention of resources the Network of Favors works fine: the collaborative actors are prioritized and granted favors, and the free riders are left with a low degree of success. That’s what is happening in this case. The peer A has 20 units of free resources, and each peer in the table is requesting 10 units of favor. A will grant D 10 units, and the last 10 units will be provided to B.
• #6: But in scenarios of low contention of resources, the surplus resources of the federation can be consumed by the free riders. In this new example, only peer F is requesting resources. Note that, for this single case, there is plenty of resources in the federation, and this increases the probability of a free rider grant more favors. We might note that peer F seems to be a free rider, because he didn’t provided any favor to A, he just consumed favors. In this scenario (low contention of resources) the peer F would grant this favor, and if he really is a free rider, this resource will not be reciprocated to A in the future.
• #7: In order to better quantify the performance of the peers in different scenarios of contention of resources, we created these two metrics: satisfaction and fairness. Satisfaction is measured by the amount of resources consumed over the amount of resources requested, which could be also referred as the probability of being answered when requesting resources. For example, if a peer requests 10 resources and he is granted these 10 resources, then, he has the best level of satisfaction possible, which is 1. Fairnes is the amount of consumed resources over the amount of donated resources. It could also be referred as the degree of reciprocity. For example, if a peer provides 10 resources and consumes only 1 unit of resources, he has a low level of fairness, 0.1. If he consumed the 10 resources it had provided, then he would have a good level of fairness, 1. Based on these 2 metrics we took the liberty of renaming the NoF, which was previously proposed to opportunistic desktop grids, to Satisfaction-Driven NoF. But why? In this NoF, every peer always provides all of its free resources to try to maximize your own satisfaction. The idea is that when he provides more resources, higher is the probability of another peer consume them. Then, the providing peer would increase his credit with the consuming peer, and this would get him to “the top of the other peers’ table”. This would increase the probability of this peer be answered when he need resources. In this case, this could be done due to the low cost of resources. But, when we come to a scenario that involves high costs in the business model, like federation of cloud providers, the costs should not be neglected. Then, in these scenarios, the fairness is the metric to be maximized, since it involves the donation of resources. Basically, what we propose is that the peers should regulate the amount of provided resources, so that the federation could avoid scenarios of very low contention of resources, which is good for free riders.
• #8: Now that we have these 2 metrics, fairness and satisfaction, we want to understand how peers perform in SD-NoF and in our proposed FD-NoF. Our simulation have the following factors, which are in red. Each simulation of a p2p network is composed by: n: the number of peers f: the amount, in percentage, of free riders steps: the number of steps C: the amount of resources each peer has D: the demand of each peer PI: the probability of being in consuming state One simplification that we did is about the time. Here it is not continuous, but discrete. In our first evaluation, we wanted to know how the set of collaborators perform in relation to free riders. Thus, all collaborators have the same Capacitys and Demands. Further, we eliminate this simplification to better understand how collaborators with higher Demand perform in comparison to another with lower demand. In the SD-NoF, a collaborator always supplies all of its free resources, in order to try to maximize his Satisfaction.
• #9: We designed the scenarios in terms of the contention of resources. What we want here is to better understand how peers perform in different levels of contention. In mathematical terms, the contention of resources (kappa) is measured by the amount of requested resources by the set of collaborators in consuming state divided by the amount of resources supplied by the set of collaborators in provider state. For example, if the consuming collaborators request in total 10 resources. If there were a set of collaborator that together could donate 20 resources, then we would have a low resource contention scenario. Note that there would be 10 exceeding resources that could be consumed by free riders. Our proposal is that the set of collaborator could regulate the amount of supplied resources to the federeation and therefore control the level of contention. Note that this set of collaborators could reduce the total amount of resources supplied from 20 to 10, and so they would have a a moderate contention of resources. Nota that here, there aren’t surplus resources that the free riders could allocate. And in a similar way they could increase even more the level of contention, what would give us a scenario of high level of contention.
• #10: We advocate that collaborators should be equipped with a control mechanism that enables them to regulate the amount of resources donated, thereby impacting the level of resource contention in the system. This should be done in such a way that high values for fairness are attained without impairing that of satisfaction.
• #14: Ao analisarmos K=0.5 e K=1, os casos onde a justiça pode ser afetada devido à sobra de recursos, podemos ver que os colaboradores apresentaram uma melhoria em ambos os cenários. Para K=0.5, os colaboradores tiveram ganho absoluto médio de 20% de justiça e e ganho relativo de 38%. Então a pergunta relativa ao trade-off seria: “vale a pena trocar 1% de satisfação por 38% de justiça?”