Abstract
Temporarily storing delay-tolerant data at peak hours and forwarding the data at off-peak hours, i.e., performing Store-and-Forward (SnF) using datacenter storage, can mitigate peak-hour congestion and exploit off-peak-hour bandwidth in inter-datacenter networks. Most prior studies considered a case where all nodes along their routing paths provided SnF options for the scheduling decision making. Intuitively, their solutions maximize the scheduling flexibility. However, the computational complexity of their solutions increases exponentially with the hop count. Meanwhile, SnF operations are generally performed on a portion of nodes rather than every node along the path. Thus, the considered case seems to be unnecessary in practice. In this paper, our studies reveal that desirable performance can be attained by involving a portion of nodes rather than all nodes along the path in the decision making. Thus, we propose a partial SnF scheduling method, which involves a portion of nodes in scheduling and introduces a network abstraction based on the involved nodes. Simulations demonstrate that the proposed method has lower complexity while achieving high performance.
Highlights
Our studies reveal that choosing the number of storage nodes involved in scheduling is a tradeoff between performance and complexity
The analytic models are presented to quantify the impact of the number of storage nodes on this tradeoff
Our key findings show that the partial SnF scheme with a larger time horizon of the temporal scheduling has the potential to outperform the full SnF scheme while maintaining lower complexity
Summary
The offpeak-hour residual bandwidth can be exploited and used for bulk data transfers This solution is called store-and-forward (SnF) and has proven to be effective in inter-DCNs [9]. The more storage nodes, the more flexible the SnF solution, and the more residual bandwidth can be utilized for bulk data transfers. In the context of inter-DC optical networks, each SnF operation will require an expensive OEO conversion, which consumes extra power and introduces control overhead [9]. This naturally inspires us to explore how to involve a portion of nodes rather than all nodes along the routing path in the decision making, thereby reducing the computational complexity
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have