Peering between Internet Service Providers (ISPs) is playing an increasingly critical role in Internet traffic exchange. As content delivery networks continue to expand, major content ISPs are increasingly opting for peering arrangements over transit services to facilitate faster exchange of traffic. The satisfaction of the ISP pair and the longevity of the peering arrangement depend on the stability and performance of these peering relationships. We introduce meta-peering, a term which refers to the set of tools needed to help and automate the ISP peering process – starting with identifying a list of ISPs that are likely to peer, writing router rules to establish BGP sessions with them, and extending the service to monitor all these sessions for notifying any major outages or peering agreement violations. In this paper, we first make a thorough analysis of recent trends in ISP peering and describe how meta-peering can be implemented by integrating some of the existing tools. We mainly focus on instrumenting the automation of the peer selection process with an aim to identifying potential peering partners and peering locations to exchange traffic. Using these direct peering links greatly reduces energy consumption as traffic takes much shorter paths to their destinations, going through reduced number of intermediary devices (e.g., routers, switches) compared to elongated transit routes, consequently reducing the environmental impact. Utilizing PeeringDB and CAIDA datasets to identify possible peering points for ISP pairs, we consider ISPs’ internal policies to generate a list of acceptable peering contracts (APCs). We design two methodologies to rank order each ISP in the APC list and offer guidance on which ones would be stable and beneficial for the potential peers. A study of more than 3,000 ISP pairs (mostly active in North America) shows that our peer selection methods can attain around 80% accuracy in predicting peering relations.

We analyze traffic exchange between Internet Service Providers (ISPs) at an Internet Exchange Point (IXP) as a non-cooperative game with ISPs as self-interested agents. Each ISP has the choice of exchanging traffic either using the shared IXP facilities, or outside the IXP – through their transit providers or private peering. We analyze the efficiency (social cost optimality) of the traffic exchange equilibrium at the IXP taking into consideration the congestion cost experienced by the ISPs at the IXP. To model both non-profit and for profit IXPs, we consider several cases, i) where the IXP does not charge any price to ISPs for the traffic exchanged (zero pricing), ii) when it charges a price that is proportional to the aggregate level of congestion at the IXP (proportional pricing), and iii) when it charges a constant price per unit traffic (constant pricing). Further, we also analyze the profit earned by the IXP under these pricing policies, under two different models of the congestion cost (delay) functions. Simulations conducted using data for actual IXPs obtained from PeeringDB demonstrate that the theoretical bounds derived for social cost and profit optimality at equilibrium (measured as the Price of Anarchy) are fairly tight, and correctly capture the performance trends against the variation of key model parameters. Further, the results show that for proportional pricing, there is an operating price range that attains near-optimal social cost and near-optimal IXP profitsimultaneously. We also demonstrate -through both theoretical analysis and simulations -that as compared to zero and constant pricing policies, proportional pricing attains better tradeoff between social cost and IXP profit, and also results in a performance that is more robust to price variations.

In this paper, we present FLASH, a framework which addresses for the first time the central AutoML problem of Combined Algorithm Selection and HyperParameter (HP) Optimization (CASH) in the context of Federated Learning (FL). To limit training cost, FLASH incrementally adapts the set of algorithms to train based on their projected loss rates, while supporting decentralized (federated) implementation of the embedded hyper-parameter optimization (HPO), model selection and loss calculation problems. We provide a theoretical analysis of the training and validation loss under FLASH, and their tradeoff with the training cost measured as the data wasted in training sub-optimal algorithms. The bounds depend on the degree of dissimilarity between the datasets of the clients, a result of FL restriction that client datasets remain private. Through extensive experimental investigation on several datasets, we evaluate three variants of FLASH, and show that FLASH performs close to centralized CASH methods.

The importance of peering in traffic exchange be-tween ISPs is rapidly increasing. With continuing growth in content delivery networks, large content ISPs are becoming increasingly inclined to exchange traffic through peering rela-tionships rather than using transit services. The stability and performance of these peering relationships dictate the satisfaction of the ISP pair and the durability of the peering. Quantification of various parameters that indicate the efficiency of peering contracts is however difficult. From the perspective of any ISP pair, there are two key decisions to be made: whether to peer or not, and at which location(s) to peer. We propose two metrics, peering willingness and peering stability, towards quantifying an ISP’s decision to peer at a location with another ISP, and the stability of that relationship. We compute these metrics using publicly available data to characterize peering relationships for different ISP pair types. We observe that peering between Content and Access ISP pairs results in the most stable and efficient relationship.

We consider a traffic peering game between Internet Service Providers (ISPs) at an Internet Exchange Point (IXP), where each ISP pair has a choice of exchanging traffic through a public IXP switch or sending the traffic through transit providers. We analyze the traffic flow efficiency (measured as social welfare) and the IXP revenue at the equilibrium of this game, as a function of the per-unit price charged by the IXP. We show that there exists a price point at which both social welfare and revenue are high, and the corresponding price-of-anarchy values can be expressed in terms of certain sublinearity measures of the inverse demand curves of the ISPs. Simulations carried out using models based on actual IXP data obtained from PeeringDB demonstrate that the theoretical bounds correctly capture the performance trends against the variation of price, and for a carefully chosen pricing point both social welfare and IXP revenue are within a factor of two of the corresponding optimal values.