Real-time multimedia applications, such as interactive gaming, live video streaming, and augmented reality, have strict latency and bitrate requirements. However, unpredictable network conditions, such as congestion and link quality, can severely degrade the quality of experience (QoE). While buffer-based mitigations cannot be applied to real-time applications due to their immediate resource needs, recent innovations in network slicing have demonstrated the feasibility of dedicating specified amounts of network resources to individual sessions in the radio access network. Encouraged by this, we propose to reserve network resources for multimedia sessions in real time according to their declared needs, thereby providing ad hoc session-level performance guarantees. Through Wi-Fi experiments and trace-driven LTE simulations, we show that such session-level resource provisioning is robust to real-time channel fluctuations and congestion externalities over the lifetime of a session. This approach, however, raises challenges: how can the network ensure that users are honest about their resource needs and optimally allocate its limited resources to users under uncertainty in future sessions' resource needs? We derive a novel multi-unit combinatorial auction (MUCA) model with a unique structure that can be exploited for fast winner determination, and yet incentivize truthful bidding, properties not simultaneously achieved in a generic MUCA but essential to making real-time session guarantees. Furthermore, since dynamic bidding in real time is challenging for end-users who are budget-constrained, we develop a reinforcement learning-based utility-maximizing strategy to distribute their budget across sessions and show that it yields high user utility.

As demand for Internet usage increases, Internet service providers (ISPs) have begun to explore pricing-based solutions to dampen data demand. However few explicitly consider the dual problem of monetizing idle network capacity at uncongested times. PopData is a recent initiative from Verizon that does so by offering supplemental discount offers (SDOs) at these times, in which users can pay a fixed fee in exchange for unlimited data in the next hour. This work is the first of its kind to assess the benefits and viability of SDOs by modeling user and ISP decisions as a game, considering both overall monthly decisions and hour-to-hour decisions throughout the month. We first use our monthly model to show that users are generally willing to accept some SDO offers, allowing the ISP to increase its revenue. We then show that users face a complex hourly decision problem as to which SDOs they should accept over their billing cycles, since they are unaware of their exact future needs or when future SDOs will be made. The ISP faces a similarly challenging problem in deciding when to offer SDOs so as to maximize its revenue, subject to users' decisions. We develop optimal decision criteria for users and ISPs to decide whether to make or accept SDO offers. Our analysis shows that both users and ISPs can benefit from these offers, and we verify this through numerical experiments on a one-week trace of 20 cellular data users. We find that ISPs can exploit user uncertainty in when future SDOs will be made to optimize its revenue.