1. Chronicle DCN Problems: Oversubscriptions, Hotspots, and Cabling Complexity

Among many other architectures, technology giants construct DCNs in a hierarchical tree topology where servers are arranged in racks as shown in Figure 1. While intra-rack communication is realized by top-of-rack edge switches (ESs) in the lower layer, inter-rack communication is established by linking ESs with aggregate switches (ASs) in the middle layer and then connecting ASs with the core switches (CSs) at the top layer. All these connections are established via uniform and fixed capacity wires/fibers. However, this hierarchical topology yields a multi-root tree where more powerful links and switches are required in the branches nearer the top layer, which makes CSs the DCN bottleneck in heavy traffic conditions. Therefore, inter-rack communication observes a throughput much lower than the actual available bandwidth [2], a.k.a. oversubscription. Moreover, measurements and analysis of real-life DCN traffic characteristics reveal that some applications generate unpredictable traffic patterns and asymmetrical traffic distributions. This unbalanced traffic is mainly caused by hotspots that contain common data required by many ongoing jobs in several DCN entities. The DCN traces also show that only 60% of the edge and core links are active at a time and utilization of 95th percentile of aggregation links is below 10% [3]. It is challenging for wired DCNs to conform with this unpredictable and unbalanced traffic because of the fixed hierarchical topology and the inflexible links with uniform capacity. Unfortunately, there is no practical way of preventing the performance degradation caused by these stochastic and unbalanced traffic conditions. Moreover, the cabling cost and complexity increase as the size of DCN scales up.

Wireless Data Centers: Towards Reconfigurable DCN Architectures

As a remedy, Wireless DCNs (WDCNs) have recently received attention to augment the aforementioned limitations of the wired DCs. Motivated by the fact that more than 95% of the entire DCN traffic is carried out by the top 10% largest flows, Skandula et. al. was first to propose mitigating hotspots by establishing wireless on-demand links in DCNs oversubscribed with such large flows [2]. Wireless technologies can provide unique DCN attributes which is summarized as follows:

• Wiring a huge number of servers is a quite demanding engineering task which exacerbates as the DCN size scales up. Since wire deployment is usually an error-prone task, necessary maintenance and modifications also consumes considerable time and efforts. On the other hand, convenient plug-and-play wireless modules can alleviate wiring cost and complexity to a great extent.
• Wireless techniques may yield a more energy and bandwidth efficient DCN by eliminating the need for switches which are typically power-hungry, error-prone, and bandwidth limited.
• WDCNs can handle oversubscription and hotspots thanks to its flexibility and reconfigurability. Given an effective hardware topology, adapting the virtual topology as per the QoS demands and traffic loads also provides a higher throughput and efficient bandwidth utilization.

Nonetheless, there exist formidable challenges in design and provisioning of WDCNs, which varies with distinct characteristics of the underlying communication technology.

B. FSO Concepts: The first proof-of-concept FSO based DC is presented in [5] where authors build a small form factor steerable FSO transceivers by using commodity components. Employing these FSO transceivers, they develop an inter-rack network solution, namely FireFly, where a ceiling mirror is preferred to establish line-of-sight FSO links between the racks. Noting that available outdoor FSO modules are bulky and expensive, authors were able to build FireFly system with low-cost commodity optical devices with high-precision and low-latency steering techniques. The experimental results show that FireFly can deliver desirable throughput and flow completion time performance comparable to Fat-Tree DCNs. Another interesting wireless DC, namely ProjecToR [c.f. Figure 2], is presented in [6] where authors use a combination of arrays of digital micromirror devices and a disco-ball shaped mirror assembly. ProjecToR is shown to provide seamless reconfigurability with high agility and fanout in comparison with Flyways, 3DBeam, and FireFly.

Virtual Topology (VirTop) Design: VirTop design is essential for leveraging the reconfigurability and flexibility of wireless DCNs to meet the dynamically changing QoS demands of network entities. In this respect, VirTop design involves many interesting and challenging open research areas such as resource allocation, scheduling, power control, beamforming, interference management, routing, etc. QoS provisioning is a critical requirement that must be guaranteed for highly diversified types of DCN traffic. Therefore, it is important to distinguish flows based on size, completion time requests, priority, etc. DC flows are typically classified as bandwidth-hungry elephant flows (EFs) or delay-sensitive mice flows (MFs). While the majority of the traffic volume is carried out by EFs, the majority of flows are typically MFs. Among the MFs, there might be higher priority critical flows with ultra-reliable and low-latency service demands. Thus, VirTop should be designed to distinguish and treat different types of flows based on their needs. For instance, MFs may experience intolerable delays if they are routed along with EFs on the same path. Hence, flow detection mechanism should be rapid, precise, and low-overhead to prevent such performance degradations.

To this end, a distributed fast, lightweight, and accurate flow detection mechanism, namely LightFD, is developed in [12]. By leveraging the TCP behavior, LightFD is designed as a module installed in the virtual-switches. However, distributed solutions are only suitable for modifiable switches with location privileges. Therefore, an alternative centralized scheme is also proposed such that a central unit reconfigures the virtual-switches by remotely communicating with switches via OpenFlow protocol. Extensive emulation results show that the accuracy of LightFD schemes outperforms the traditional methods. Moreover, the detection speed of the distributed and centralized schemes are in the order of milliseconds and ten milliseconds, respectively.

The complexity of VirTop design increases not only with the size but also with the massive and diverse amount of traffic generated across the DCN. By bundling and treating the same class of flows destined to a common destination, flow grooming (FG) can increase the bandwidth utilization and relax the complexity of the network management. In this direction, a three- step FG (3SFG) policy is developed for a two-layer wireless DCN topology in [13] where leaf and spine layers respectively comprise of ESs and CSs, which are interconnected via WDM-FSO links. In 3SFG, servers first groom MFs destined to the same server. Then, servers re-groom the flows in the first step intended to the same rack. Finally, ESs performs a final grooming to create rack-to-rack (R2R) jumbo flows. Since MFs forms the majority of the generated flows, 3SFG calculates predetermined R2R routing paths whose capacity is calculated based on the flow arrival statistics, flow size, and required flow completion duration. In this manner, a significant portion of flows is handled with a very low complexity by ensuring the QoS demands. The remaining wavelengths (i.e., channels or RBs) are exploited by EFs which are transferred over server-to-server (S2S) express routes. By routing MFs and EFs on disjoint paths, 3SFG prevents highly probable congestion occasions as delay sensitive MFs can be affected by bandwidth-hungry EFs easily. On top of this, 3SDG immediately serves mission critical flows over available S2S routes by interrupting the ongoing EFs. Noting the small size of the MFs and available large capacity, such an impulsive interruption does not harm the EF performance distinguishably. Numerical results show that 3SFG can significantly improve the throughput and flow completion time.

Security and Privacy: Nowadays it is a common practice for companies to rent a portion of mega DCNs in order to prevent maintaining and upgrading costs of a self-owned enterprise network. In this case, it is critical to isolate wirelessly transmitted data from unintended nodes in order to prevent security and privacy risks. Fortunately, the directivity and limited penetration of high-speed wireless links make it possible to partition the DCN via physical obstacles. Such an approach may enhance the immunity to eavesdropping, allow to develop low-overhead security protocols, and hence increase overall networking performance. Security and privacy are in the center of skepticism of the cloud service providers toward the WDCNs as they are legally responsible for protecting the sensitive information.

Conclusions

In this newsletter, we provide an overview of wireless DCNs for embracing the emerging information and communication technologies that require storage and processing of the highly diverse massive amount of data. After briefing and comparing the virtues and drawbacks of potential high-speed wireless technologies, we have presented advances and challenges of implementing WDCNs from both theoretical and practical points of views. Following the presentation of ongoing research efforts focused on PhyTop design, we also discussed overlooked aspects such as VirTop design, flow detection, flow grooming, and load balancing.