We present a decade of evolution and production experience with Jupiter datacenter network fabrics. In this period Jupiter has delivered 5x higher speed and capacity, 30% reduction in capex, 41% reduction in power, incremental deployment and technology refresh all while serving live production traffic. A key enabler for these improvements is evolving Jupiter from a Clos to a direct-connect topology among the machine aggregation blocks. Critical architectural changes for this include: A datacenter interconnection layer employing Micro-Electro-Mechanical Systems (MEMS) based Optical Circuit Switches (OCSes) to enable dynamic topology reconfiguration, centralized Software-Defined Networking (SDN) control for traffic engineering, and automated network operations for incremental capacity delivery and topology engineering. We show that the combination of traffic and topology engineering on direct-connect fabrics achieves similar throughput as Clos fabrics for our production traffic patterns. We also optimize for path lengths: 60% of the traffic takes direct path from source to destination aggregation blocks, while the remaining transits one additional block, achieving an average block-level path length of 1.4 in our fleet today. OCS also achieves 3x faster fabric reconfiguration compared to pre-evolution Clos fabrics that used a patch panel based interconnect.

Data Center topologies employ multiple paths among servers to deliver scalable, cost-effective network capacity. The simplest and the most widely deployed approach for load balancing among these paths, Equal Cost Multipath (ECMP), hashes flows among the shortest paths toward a destination. ECMP leverages uniform hashing of balanced flow sizes to achieve fairness and good load balancing in data centers. However, we show that ECMP further assumes a balanced, regular, and fault-free topology, which are invalid assumptions in practice that can lead to substantial performance degradation and, worse, variation in flow bandwidths even for same size flows.

We present the design, implementation, and evaluation of B4, a private WAN connecting Google's data centers across the planet. B4 has a number of unique characteristics: i) massive bandwidth requirements deployed to a modest number of sites, ii) elastic traffic demand that seeks to maximize average bandwidth, and iii) full control over the edge servers and network, which enables rate limiting and demand measurement at the edge.

The broadcast nature of existing wireless LANs can produce excessive radio coverage which is often undesirable for applications like wireless IPTV or location-based service. To address this issue, we utilize the characteristics of signal attenuation common to all wireless communications and implement apoint-castingservice1(PCast) over existing wireless networks that can significantly restrict the coverage for transmissions. The principle of PCast is to split and encode the downlink traffic and deliver it to the client via a few surrounding access points (APs). Hence, an unintended client cannot decode a packet without all its fragments. Equivalently, the service area is bounded within the intersection of communication coverage of all APs. The result coverage area can be further reduced via rate and power control at the APs. In this paper, we present our design and implementation of the PCast service on top of an 802.11a wireless LAN. Experiments conducted in a lab environment show that PCast can bound the service region to be within 2 meters around a client, while incurring jitter of less than 1ms, and bandwidth capacity loss of less than 10%.

In this paper, we present a high fidelity and efficient emulation framework called TWINE, which combines the accuracy and realism of emulated and physical networks and the scalability and repeatability of simulation in an integrated testbed, for evaluation of real protocols and applications. Our measurements show that the TWINE emulation kernel has a memory footprint of less than 100KB, and occupies no more than 3.5 % CPU cycles. Thanks to such small overhead and the accurate modelling of physical layer events(at microseconds level), application throughput measured in TWINE is within 5 % of the measured throughput from an equivalent physical wireless LAN. A single commodity PC in TWINE can emulate at least four wireless hosts or simulate sixty nodes in real time at microseconds granularity. This paper also illustrates TWINE’s novel capabilities via two case studies: a protocol to maintain fairness in mesh networks and an adaptive streaming media application operating in heterogeneous wireless networks. The results from the case studies clearly show the benefit of the TWINE evaluation methodology, by identifying a mismatch between the performance of the protocol or application based on actual user experience versus its performance as measured using traditional network performance metrics such as application throughput. 1.

Recent years have seen the growing popularity of multi-rate wireless network devices (e.g., 802.11a cards) that can exploit variations in channel conditions and improve overall network throughput. Concurrently, rate adaptation schemes have been developed that selectively increase data transmissions on a link when it offers good channel quality. In this paper, we propose a Medium Access Diversity (MAD) scheme that leverages the benefits of rate adaptation schemes by aggressively exploiting multiuser diversity. The basic mechanism of MAD is to obtain instantaneous channel condition information from multiple receivers and selectively transmit data to a receiver that improves the overall throughput of the network, while maintaining temporal fairness among multiple data flows. We identify and address the challenges in the design and implementation of MAD's three phases: channel probing, data transmission, and receiver scheduling. We also use analytical models to examine the tradeoff between network performance improvement and overhead of channel probing, and derive an asymptotic performance bound for the receiver scheduling algorithms used by MAD. Results from the analysis and the extensive simulations demonstrate that, on average, MAD can improve the overall throughput of IEEE 802.11 wireless LANs by 50% as compared with the best existing rate adaptation scheme.

Discrete event network simulators have emerged as popular tools for verification and performance evaluation for various wireless networks. Nevertheless, the desire to model such networks at high fidelity implies high computational costs, prohibiting most researchers from simulating wireless networks with thousands of nodes. There have been attempts on performance optimizations for large-scale wireless network simulation, but they have not appropriately modeled accumulation of weak interference, thereby suffering inaccuracies which may be magnified by upper layer protocols. This paper presents analysis of the effects of common optimization techniques for large-scale wireless network simulation on the overall network performance and also proposes modifications and novel techniques that introduce only limited inaccuracies or no additional inaccuracy at all. The study quantifies the effects of those optimizations on the simulation results for given thresholds and network parameters, and also identifies thresholds tolerable to most network studies. The experimental results show that these optimizations can improve the runtime performance of an already efficient wireless network simulator substantially, by a factor of up to 55 for wireless networks with 3200 nodes without compromising accuracy of the simulation results.

This paper presents ARC (Adaptive Range Control), a communication range control mechanism using directional antennas to be implemented across multiple layers. ARC uses directional reception for range control rather than directional transmission such that extended communication links do not increase interference to other ongoing communications. It adaptively controls the communication range by estimating dynamically changing local network density based on the transmission activities around each network node. The experimental results using simulation with detailed physical layer, IEEE 802.11 DCF MAC, and AODV protocol models have shown the successful adaptation of communication range with ARC for varied network densities and traffic loads. ARC improves the packet delivery ratio by a factor of 9 at the maximum for sparse networks while it maintains the increased network capacity for dense networks.