跳转至

cISP: A Speed-of-Light Internet Service Provider

这篇论文主要探讨了如何构建一个传输速度接近光速的广域网(Wide-Area Network), 以解决当前基于光纤的互联网延迟过高的问题.

(1) Motivation

  • 现状: 互联网延迟远高于物理极限. 当前的延迟通常是光速延迟(c-latency)的 37倍 甚至更多

  • 原因:

    1. 介质限制: 光在光纤中的传播速度仅为真空光速的约 2/3
    2. 路径迂回: 光纤管道通常沿着铁路或公路铺设, 路径非常曲折
    3. 协议低效: 传输层和应用层的往返(RTT)开销

  • 目标: 设计一个名为 cISP 的互联网服务提供商, 利用微波(Microwave)技术, 使数据传输速度接近真空中光速

(2) CISP Design

  • 技术基础: 使用微波(MW)在塔与塔之间进行视距(Line-of-Sight)传输. 微波在空气中的速度几乎等于真空光速

  • 混合架构:

    • cISP 网络: 用于传输对延迟敏感的高价值小流量(如游戏, 即时通讯, 算法交易)
    • 传统光纤: 继续用于传输对带宽要求高但对延迟不敏感的大流量(如视频下载)

  • 拓扑构建:

    • 节点选择: 选取主要人口中心(如美国前200大城市合并后的120个节点)
    • 链路设计: 基于地理地形数据(NASA数据)计算塔与塔之间的可视性(Fresnel zone), 克服地球曲率和障碍物
    • 容量扩展: 单个微波链路带宽有限(约1Gbps). 通过构建并行的塔链来扩展带宽

(3) Feasibility & Challenges

  • 天气影响: 降水(雨/雪)会衰减微波信号
    • 应对: 论文通过一年的气象数据模拟发现, 即使在恶劣天气下, cISP 的延迟(99分位)仍优于光纤
  • 成本分析:
    • 构建全美 cISP 网络的成本估算约为 0.81$/GB, 这与 CDN 的成本相当, 在经济上是可行的
  • 物理限制:
    • 塔的可用性: 考虑了塔的高度, 位置租赁等现实约束
    • 视距限制: 典型的跳跃距离(Hop distance)限制在 60-100km

(4) Use Cases

  • 算法交易(HFT): 已经在用, 但 cISP 将其扩展到更广的范围
  • 在线游戏: 提供更公平, 响应更快的游戏体验
  • CDN 加速: 用于源站与边缘节点之间的快速数据同步

Introduction

User experience in many interactive network applications depends crucially on achieving low latency. Even seemingly small increases in latency can negatively impact user experience, and, subsequently, revenue for service providers: Google, for example, quantified the impact of an additional 400 ms of latency in search results as 0.7% fewer searches per user [18]. Further, wide-area latency is often the bottleneck, as Facebook’s analysis of over a million requests found [21]. Indeed, content delivery networks (CDNs) present latency reduction and its associated increase in conversion rates as one of the key value propositions of their services, citing, e.g., a 1% loss in sales per 100 ms of latency for Amazon [2]. In spite of the significant impact of latency on performance and user experience, the Internet is not designed to treat low latency as a primary objective. This is the problem we address: reducing latencies over the Internet to the lowest possible.

The best achievable latency between two points along the surface of the Earth is determined by their geodesic distance divided by the speed of light, c. Latencies over the Internet, however, are usually much larger than this minimal “c-latency”: recent measurement work found that fetching even small amounts of data over the Internet typically takes 37× longer than the c-latency, and often, more than 100× longer [16]. This delay comes from the many round-trips between the communicating endpoints, due to inefficiencies in the transport and application layer protocols, and from each round-trip itself taking 3-4× longer than the c-latency [16]. Given the approximately multiplicative role of network roundtrip times (RTTs) when bandwidth is not the main bottleneck, eliminating inflation in Internet RTTs can potentially translate to up to 3-4× speedup, even without any protocol changes. Further, as protocol stack improvements get closer to their ideal efficiency of one RTT for small amounts of data, the RTT becomes the singular network bottleneck. Similarly, for well-designed applications dependent on persistent connectivity between two fixed locations, such as gaming, nothing other than resolving this 3-4× “infrastructural inefficiency” can improve latency substantially.

Thus, beyond the networking research community’s focus on protocol efficiency, reducing the Internet infrastructure’s latency inflation is the next frontier in research on latency. While academic research has typically treated infrastructural latency inflation as an unresolvable given, we argue that this is a high-value opportunity, and is much more tractable than may be evident at first.

What are the root causes of the Internet’s infrastructural inefficiency, and how do we ameliorate them? Large latencies are partly explained by poor use of existing fiber infrastructure: two communicating sites often use a longer, indirect route because their service providers do not peer over the shortest fiber connectivity between their locations. We find, nevertheless, that even latency-optimal use of all known fiber conduits, computed via shortest paths in the InterTubes dataset [34], would leave us 1.98× away from c-latency [17]. This gap stems from the speed of light in fiber being ∼ 2/3 c, and the unavoidable circuitousness of fiber routes due to topographic and economic constraints of buried conduits.

We thus explore the design of cISP, an Internet Service Provider that provides nearly speed-of-light latency by exploiting wireless electromagnetic transmissions, which can be realized with point-to-point microwave antennas mounted on towers. This approach holds promise for overcoming both the aforementioned shortcomings fundamental to today’s fiber-based networks: the transmission speed in air is essentially equal to c, and the richness of existing tower infrastructure makes more direct paths possible. Nevertheless, it also presents several new challenges, including:

• overcoming numerous practical constraints, such as tower availability, line-of-sight requirements, and the impact of weather on performance;

• coping with limited wireless bandwidth;

• solving a large-scale cost-optimal network design problem, which is NP-hard; and

• addressing switching and queuing delays, which are more prominent with the smaller propagation delays.

To meet these challenges, we propose a hybrid design that augments the Internet’s fiber connectivity with nearly straight-line wireless links. These low-latency links are used judiciously where they provide the maximum latency benefit, and only for the high-impact but small proportion, in terms of bytes, of Internet traffic that is latency-sensitive. We design a simple heuristic that achieves near-optimal results for the network design problem. Our approach is flexible and enables network design for a variety of deployment scenarios; in particular, we show that cISP’s design for interconnecting large population centers in the contiguous U.S. and Europe can achieve mean latencies as low as 1.05× c-latency at a cost of under $1 per gigabyte (GB). We show through simulation that such networks can be operated at high utilization without excessive queuing.

To address the practical concerns, we use fine-grained geographic data and the relevant physical constraints to determine where the needed wireless connectivity would be feasible to deploy, and assess our design under a variety of scenarios with respect to budget, tower height and availability, antenna range, and traffic matrices. We also use a year’s worth of meteorological data to assess the network’s performance during weather disturbances, showing that most of cISP’s latency benefits remain intact throughout the year. Our weather simulation and an animation showing how the hybrid network evolves from mostly-fiber to mostly-wireless with increasing budget are available online; see [25] and [26].

But is it feasible to use microwave hardware for low latency in practice? To answer this question, we rented virtual machines in the CME data center in Chicago and the Equinix data center in New Jersey, and, on Saturdays, were given access at these data centers to one of the fastest microwave networks spanning the Chicago – New Jersey algorithmic trading corridor. Experiments conducted on this network show that it successfully operates at a speed extremely close to the speed of light, and that losses can be effectively handled by extremely lightweight forward error correction (FEC). We complement these findings by analyzing real trading data, revealing the minimum latency between the data centers and showing that the network is available in varied weather conditions.

Finally, we explore the application-level benefits for Web browsing and gaming, and present estimates showing that the utility of cISP vastly exceeds its cost, even for web sites already using CDNs to reduce latency.

许多交互式网络应用的用户体验关键取决于能否实现低延迟. 即便是微小的延迟增加, 也可能对用户体验乃至服务提供商的营收产生负面影响: 例如, Google 量化了搜索结果中额外增加 400 毫秒延迟的影响, 发现这导致每位用户的搜索量减少了 0.7% [18]. 此外, 正如 Facebook 对超过一百万次请求的分析所发现的那样, 广域网延迟往往是瓶颈所在 [21].

事实上, 内容分发网络(CDN)将降低延迟及其带来的转化率提升作为其服务的核心价值主张之一, 例如引用数据指出, 亚马逊每增加 100 毫秒的延迟会导致销售额损失 1% [2]. 尽管延迟对性能和用户体验有着显著影响, 但互联网的设计初衷并未将低延迟视为首要目标. 这正是我们所要解决的问题: 将互联网上的延迟降至最低限度.

地球表面两点之间可达到的最佳延迟, 取决于两点间的测地距离(geodesic distance)除以光速.

然而, 互联网上的实际延迟通常远高于这一最小的"-延迟(c-latency)": 近期的测量研究发现, 通过互联网获取少量数据通常需要花费 "c-延迟" 的 37 倍时间, 甚至经常超过 100 倍 [16].

这种延迟源于:

  1. 通信端点之间多次的往返交互(由于传输层和应用层协议的低效)
  2. 每次往返本身所需时间是 "c-延迟" 的 3-4 倍 [16]

鉴于在带宽非主要瓶颈时网络往返时间(RTT)所起的近似乘数效应, 消除互联网 RTT 的膨胀(inflation)有望带来高达 3-4 倍的加速, 且无需更改任何协议.

此外, 随着协议栈的改进逐渐接近处理少量数据只需一次 RTT 的理想效率, RTT 将成为唯一的网络瓶颈. 同样, 对于依赖两处固定位置间持久连接的精心设计的应用(如网络游戏), 除了解决这 3-4 倍的"基础设施低效"问题外, 别无他法能显著改善延迟.

因此, 除了网络研究界对协议效率的关注之外, 减少互联网基础设施的延迟膨胀是延迟研究的下一个前沿领域. 尽管学术研究通常将基础设施的延迟膨胀视为不可解决的既定事实, 但我们要论证的是, 这是一个高价值的机遇, 且比乍看之下更易于处理.

互联网基础设施低效的根本原因是什么? 我们又该如何改善它们?

高延迟部分归因于对现有光纤基础设施的利用不足: 两个通信站点往往使用较长的间接路由, 因为它们的服务提供商并未在其位置之间最短的光纤连接上进行对等互连(peering).

尽管如此, 我们发现, 即使通过 InterTubes 数据集 [34] 中的最短路径计算, 以延迟最优的方式利用所有已知的光纤管道, 我们距离 "c-延迟" 仍有 1.98 倍的差距 [17].

这一差距源于光在光纤中的传播速度约为 2/3 c, 以及由于埋设管道受地形和经济约束而不可避免的路径迂回

因此, 我们探索了 cISP 的设计, 这是一种利用无线电磁传输提供近乎光速延迟的互联网服务提供商(ISP), 该方案可通过安装在塔上的点对点微波天线来实现. 这种方法有望克服当今基于光纤的网络所固有的上述两个缺陷: 空气中的传播速度本质上等于 c, 且现有的丰富塔架基础设施使得更直接的路径成为可能.

然而, 这也带来了若干新的挑战, 包括:

  • 克服诸多实际限制, 如塔架可用性, 视距(Line-of-Sight)要求以及天气对性能的影响
  • 应对有限的无线带宽
  • 解决大规模成本最优网络设计问题(属 NP-hard 问题)
  • 解决交换和排队延迟, 这在传播延迟较小的情况下显得更为突出

为应对这些挑战, 我们提出了一种混合设计, 即利用 近乎直线的无线链路来增强互联网的光纤连接

这些低延迟链路被审慎地用于能提供最大延迟收益的地方, 且仅用于传输那些对延迟敏感, 影响重大但字节占比很小的互联网流量.

我们设计了一种简单的启发式算法, 可针对网络设计问题取得近乎最优的结果. 我们的方法具有灵活性, 支持针对多种部署场景的网络设计; 特别是, 我们证明了 cISP 用于互联美国本土和欧洲主要人口中心的设计, 能够以每千兆字节(GB)低于 1 美元的成本, 实现低至 1.05 倍 c-延迟的平均延迟. 我们通过仿真表明, 此类网络可以在高利用率下运行, 而不会产生过度的排队现象.

为解决实际顾虑, 我们利用细粒度的地理数据及相关物理约束, 来确定所需无线连接的可行部署位置, 并针对预算, 塔高与可用性, 天线覆盖范围及流量矩阵等多种场景评估我们的设计. 我们还使用了一整年的气象数据来评估网络在恶劣天气下的性能, 结果显示 cISP 的大部分延迟优势在全年内都能得以保持. 我们的天气模拟和展示混合网络如何随着预算增加从主要基于光纤演变为主要基于无线的动画均可在线获取; 参见 [25] 和 [26].

但在实践中, 使用微波硬件实现低延迟是否可行?

为回答这一问题, 我们在芝加哥的 CME 数据中心和新泽西的 Equinix 数据中心租用了虚拟机, 并在周六获准通过这些数据中心访问跨越芝加哥--新泽西算法交易走廊的最快微波网络之一. 在该网络上进行的实验表明, 其运行速度极其接近光速, 且极其轻量级的前向纠错(FEC)技术可有效处理丢包问题. 我们通过分析真实交易数据补充了这些发现, 揭示了数据中心间的最小延迟, 并表明该网络在各种天气条件下均可用.

最后, 我们探讨了其在网页浏览和游戏方面的应用层收益, 并提出了估算结果, 显示 cISP 的效用远超其成本, 即便是对于已经使用 CDN 降低延迟的网站也是如此.

看到这的第一直觉

"在我看来, 本文的创新点一句话概括就是用 mmWave "代替"传统以太网, 这样更靠近"理论上限C", 但问题是, 这不就是5G吗? 我很清楚的记得5G的中高频段就是使用 mmWave 的"

5G mmWave 和 cISP 在物理层利用的特性是一致的: 电磁波在空气中的传播速度接近光速 c, 而光在光纤中的速度约为 2/3 c

但是, 说 cISP "就是 5G" 并不准确. cISP 的核心创新不在于"发现微波比光纤快", 而在于 Architecture and Topology) 的根本不同

(1) 定位不同: 骨干网 vs. 接入网

  • 5G (Access Network):
    • 目标: 解决"最后一公里"问题, 连接UE和gNB
    • 瓶颈: 5G 的空口延迟确实很低(uRLLC 可达 1ms 级别), 但数据一旦进入基站, 回传(Backhaul)通常还是走光纤
  • cISP (Long-haul Transit):
    • 目标: 解决"长途传输"问题, 连接城市与城市(City-to-City)
    • 创新: 它是要替代光纤骨干网. cISP 试图建立一个横跨大陆的无线接力网络

(2) 拓扑与链路技术: P2P vs. P2MP

  • 5G mmWave:

    • 模式: 点对多点(Point-to-Multipoint)
    • 挑战: 需要处理移动性, 波束赋形(Beamforming), 复杂的多址接入(Multi-access), 调度
    • 距离: mmWave 基站覆盖半径通常仅为几百米

  • cISP:

    • 模式: 点对点(Point-to-Point, P2P)
    • 技术: 使用的是高增益定向天线(如微波大锅), 类似于高频交易(HFT)目前使用的专用微波链路
    • 距离:
      • 论文中提到的 Hop distance 为 40km - 100km
      • 这通常需要使用比 5G mmWave 更低的频段(如 6-11 GHz 微波)或者高功率的 E-band 才能实现, 而不是 5G FR2 的那种短距覆盖

Tech Background

At the highest level, our approach involves using free-space communication between transmitters mounted at a suitable height, e.g., using dedicated towers or existing buildings, and separated from each other by at most a certain limiting distance. Network links longer than this range require a series of such transmitters. Typically, even after accounting for terrain, such a network link can be built close to the shortest path on the Earth’s surface between the two end points. Further, the speed of light in air is essentially the same as that in vacuum, c. These properties make our approach attractive for the design of (nearly) c-latency networks.

Technology choices. Several physical layer technologies are amenable for use in our design, including free-space optics (FSO), microwave (MW), and millimeter wave (MMW). At present, we believe MW provides the best combination of range, resilience, throughput, and cost. Future advances in any of these technologies, however, can be easily rolled into our design, and can only improve our cost-benefit analysis.

While hollow fiber [31] could, in the future, also provide clatency, it would still suffer from the circuitousness of today’s fiber conduits. Low Earth orbit satellite networks, as are being currently deployed, could also help, although they currently incur substantially higher latency than cISP (§9).

Switching latency. While long-haul MW networks have existed since the 1940s [10], their use in high-frequency trading starting within the last 10 years [55] has driven innovation in radios so that each MW retransmission only takes a few µs. Thus, even wide-area links with many retransmissions incur negligible switching latency. As an example, the HFT industry operates a MW relay between Chicago and New Jersey comprising ≈ 20 line-of-sight links that operates within 1% of c-latency end-to-end at the application layer [58].

Packet loss. Loss occurs for several reasons, including weather disruption and intermittent multi-path fading, especially over bodies of water. In §5.1, using a year’s worth of weather data, we analyze the impact of diverting traffic to alternate (fiber or MW) routes during inclement weather. Our active experiments on a microwave network also show that losses experienced could be handled with lightweight forward error correction (FEC).

Spectrum and licensing. We propose the use of MW communication in the 6-18 GHz frequency range. These frequencies are not very crowded, and licensing is generally not very competitive, except at 6 GHz in cities, and along certain routes, like the above mentioned HFT corridor. The licenses are given on a first-come, first-served basis, recorded in a public database, they protect against the deployment of other links that would interfere with licensed links.

Line-of-sight & range. Successive MW towers need lineof-sight visibility, accounting for the Earth’s curvature, terrain, trees, buildings and other obstructions, and atmospheric refraction. Attenuation also limits range. A maximum range of around 100 km is practicable, but we show results with maximum allowed range varying between 60-100 km (§5.2).

Bandwidth. Between any two towers, using very efficient encoding (256 QAM or higher), wide frequency channels, and radio multiplexing, a data rate of about 1 Gbps is achievable [45]. This bandwidth is vastly smaller than for fiber, and necessitates a hybrid design using fiber and MW.

Geographic coverage. Connecting individual homes directly to such a MW network would be cost-prohibitive. To maximize cost-efficiency, we focus on long-haul connectivity, with the last mile being traditional fiber. At short distances, fiber’s circuitousness and refraction are small overheads.

Cost model. We rely on cost estimates in recent work [55] and based on our conversations with industry participants involved in equipment manufacturing and link provisioning. The cost of installing a bidirectional MW link, on existing towers, is approximately $75K ($150K) for 500 Mbps (1 Gbps) bandwidth. The average cost for building a new tower is $100K, with wide variation by terrain and across cities and rural areas. Any additional towers needed to augment bandwidth for particular links incur this “new tower” cost. The operational costs comprise several elements, including management and personnel, but the dominant operational expense, by far, is tower rent: $25−50K per year per tower. We estimate cost per GB by amortizing the sum of building costs and operational costs over 5 years.

Note that the deployment and operational costs can vary substantially based on the deployment model. For example, imagine that a company like American Tower [7], which has a substantial tower presence across the US (see Fig. 14 in Appendix D), deploys cISP. In such a scenario, not only would the cost of bandwidth augmentation be negligible, but also the cost of maintaining the towers would be drastically reduced. We consider both conservative and optimistic deployment models and conduct an in-depth cost-analysis in this work.

从宏观层面来看, 我们的方法涉及利用自由空间通信技术, 通过安装在适当高度(例如专用塔架或现有建筑物上)且间距不超过特定极限距离的发射机进行数据传输. 网络链路若超过此范围, 则需由一系列此类发射机级联而成. 即便将地形因素纳入考量, 此类网络链路通常仍能紧贴两端点间地球表面的最短路径构建. 此外, 光在空气中的传播速度本质上与真空光速 c 相同. 这些特性使得我们的方法在设计(近乎)c-延迟网络方面极具吸引力.

技术选择. 多种物理层技术均适用于我们的设计, 包括自由空间光通信(FSO), 微波(MW)和毫米波(MMW). 目前, 我们认为微波在传输距离, 韧性, 吞吐量和成本方面提供了最佳的综合性能. 未来任何此类技术的进步均可轻松融入我们的设计, 且只会进一步改善我们的成本效益分析.

尽管空芯光纤 [31] 在未来也可能实现 c-延迟, 但其仍将受制于当今光纤管道的迂回性问题. 目前正在部署的低地球轨道(LEO)卫星网络虽有助益, 但其产生的延迟目前仍显著高于 cISP(详见 §9)

交换延迟. 尽管长途微波网络自 20 世纪 40 年代 [10] 即已存在, 但过去十年间高频交易(HFT) [55] 的应用推动了无线电设备的创新, 使得每次微波重传仅需数微秒(µs). 因此, 即便是包含多次重传的广域链路, 其产生的交换延迟也可忽略不计. 作为佐证, HFT 行业运营着一条连接芝加哥和新泽西的微波中继线路, 包含约 20 个视距链路, 其端到端应用层延迟控制在 c-延迟的 1% 以内 [58].

丢包. 丢包的产生有多种原因, 包括天气干扰和间歇性的多径衰落, 特别是在跨越水体时. 在 §5.1 中, 我们利用一年的气象数据, 分析了在恶劣天气期间将流量分流至备用(光纤或微波)路径的影响. 我们在微波网络上的主动实验也表明, 所经历的丢包可以通过轻量级的前向纠错(FEC)技术有效处理.

频谱与许可. 我们建议使用 6-18 GHz 频段的微波通信. 该频段并不十分拥挤, 且除了城市中的 6 GHz 频段以及某些特定路线(如上述 HFT 走廊)外, 频谱许可的竞争通常并不激烈. 许可证的发放遵循"先到先得"原则, 并记录在公共数据库中, 以防止其他链路的部署对已获许可的链路造成干扰.

视距与覆盖范围. 连续的微波塔之间需要保持视距可见性, 这需要综合考量地球曲率, 地形, 树木, 建筑物及其他障碍物, 以及大气折射的影响. 信号衰减也限制了传输范围. 约 100 公里的最大传输距离在实践中是可行的, 而在 §5.2 中, 我们展示了最大允许距离在 60-100 公里之间变化时的结果.

带宽. 在任意两座塔之间, 通过使用高效编码(256 QAM 或更高), 宽频信道及无线电复用技术, 可实现约 1 Gbps 的数据速率 [45]. 该带宽远小于光纤带宽, 因此必须采用结合光纤与微波的混合设计.

地理覆盖. 将个别家庭直接连接到此类微波网络在成本上是不可接受的. 为最大化成本效益, 我们专注于长途连接, 而"最后一公里"则沿用传统光纤. 在短距离传输中, 光纤的路径迂回和折射带来的开销相对较小.

成本模型. 我们依据近期研究 [55] 中的成本估算, 并结合了与设备制造及链路供应行业参与者的交流. 在现有塔架上安装双向微波链路的成本约为 7.5 万美元(500 Mbps)至 15 万美元(1 Gbps). 新建一座塔的平均成本为 10 万美元, 具体取决于地形及城乡差异. 任何为特定链路增加带宽所需的额外塔架均按此"新塔"成本计算. 运营成本包含管理与人员等多个要素, 但其中最大的运营开支无疑是塔架租金: 每座塔每年约 2.5 万至 5 万美元. 我们通过将建设成本与 5 年的运营成本摊销来估算每 GB 的成本.

需要注意的是, 部署和运营成本会根据部署模式的不同而有显著差异. 例如, 设想像 American Tower [7] 这样在美国拥有大量塔架资源的公司(见附录 D 图 14)来部署 cISP. 在此类场景下, 不仅带宽扩容的成本微乎其微, 塔架维护成本也将大幅降低. 我们在本工作中同时考虑了保守和乐观的部署模式, 并进行了深入的成本分析.

cISP Design

alt text

(1) Hybrid Architecture

双层网络设计: cISP 并非完全取代现有互联网, 而是作为一个覆盖网络(Overlay)存在

  • cISP 层(微波): 专为低延迟, 低带宽流量服务(如游戏状态更新, VoIP 信令, 算法交易)
  • 光纤层(传统): 继续承载高带宽, 对延迟不敏感的流量(如视频流, 大文件下载)

流量分流: 系统会根据应用需求, 智能地将流量路由到光纤或 cISP 微波链路

(2) Physical Topology Construction

  • 节点选择: 选取主要人口中心(如美国人口最多的城市聚类)作为核心节点
  • 链路可行性分析:
    • 利用 NASA 的航天飞机雷达地形测绘任务(SRTM)数据, 结合地球曲率和菲涅尔区(Fresnel zone)要求, 计算塔与塔之间是否存在视距(Line-of-Sight, LoS)
    • 设定了物理约束: 如塔高(假设 60-100米), 最大跳距(Hop distance, 通常限制在 60-100km 以内以保证信号质量)

(3) Network Design Problem

  • 核心挑战: 在有限的预算(Budget)下, 如何选择建设哪些微波链路, 以最大化降低全网的平均延迟? 这是一个 NP-难(NP-hard) 的成本受限网络设计问题
  • 优化目标: 最小化所有流量对(Traffic Pairs)的加权延迟总和
  • 启发式算法 (Heuristic Algorithm): 由于计算最优解太慢, 论文设计了一种贪心策略
    • 迭代添加: 每次选择那些能带来最大"延迟收益/成本比(Benefit-per-cost)"的链路加入网络
    • 修剪: 去除冗余或收益过低的链路

(4) Capacity Planning

  • 带宽瓶颈: 单个微波链路的带宽(~1Gbps)远低于光纤
  • 扩展策略:
    • 平行链路: 在高流量路径上, 通过建设平行的塔链(Parallel tower chains)或在同一塔上安装多组天线来叠加带宽
    • 按需分配: 只有必须"快"的数据才走微波, 从而节省宝贵的微波带宽资源

(5) Routing & Interface

  • 最短路路由: 在 cISP 网络内部, 流量基于物理距离(延迟)进行最短路径转发
  • 与现有互联网集成: cISP 的路由器作为自治系统(AS)与传统 ISP 进行 BGP 互联, 向外宣告其覆盖的前缀, 吸引特定流量进入

Networking research has made significant progress in measuring latency, as well as improving it through transport, routing, and application-layer changes. However, the underlying infrastructural latency has received little attention and has been assumed to be a given. This work proposes a speed-of-light ISP, demonstrating that improvements are indeed possible.

There are several ongoing Internet infrastructure efforts, including X moonshot factory’s project Taara [90], Facebook connectivity’s Magma [36], Rural Access [37], Terragraph [38], and the satellite Internet push by Starlink [81], Kuiper [54], Telesat [86], and others. Project Taara consists of networks under deployment in India and Africa, based on free-space optics, and described as “Expanding global access to fast, affordable internet with beams of light”. While Facebook’s Magma and Rural Access aim to extend connectivity to rural areas by offering a software, hardware, business model, and policy framework, Terragraph aims to extend lastmile connectivity to poorly connected urban and suburbans areas by leveraging short millimeter-wave hops. Free-space networks of this type will likely become more commonplace in the future, and these works are further evidence that many of the concerns with line-of-sight networking can indeed be addressed with careful planning. Further, cISP’s design approach is flexible enough to incorporate a variety of media (fiber, MW, MMW, free-space optics, etc.) as the technology landscape changes.

“New Space” satellite networks: While low-Earth orbit (LEO) satellite networks can reduce long-distance latency [12, 44, 52], current deployments are more targeted at last-mile connectivity than long haul [15]. Starlink recently claimed to offer last-mile round-trip latency of 31 ms [82], more than 3.8× the latency estimated in prior simulations [12], showing that the service is not yet latency optimized.

Despite the apparent differences in objectives — long haul latency for cISP and last-mile connectivity for LEO networks — it is useful to coarsely assess how the costs may compare. Starlink, for example, offers uncapped connectivity at $99/month [78]. At an average household consumption of 273.5 GB [35], this translates to $0.36/GB For cISP, if an incumbent like American Tower were to deploy it, the cost could be as low as $0.33/GB, as shown in Fig. 3c. Thus, a network with costs comparable to cISP (in a per-bit sense; cISP is more than an order of magnitude cheaper in absolute cost, and has commensurately lower bandwidth) is concurrently being deployed, albeit with different goals.

To the best of our knowledge, the only efforts primarily focused on wide-area latency reduction through infrastructural improvements are in niches, such as the point-to-point links for financial markets [55], and isolated submarine cable projects aimed at shortening specific Internet routes [67,69].

(1) 学术界研究现状

  • 现有重点: 主要集中在传输层, 路由层和应用层的协议优化, 以及延迟测量
  • 研究空白: 底层的"基础设施延迟(Infrastructural Latency)"往往被视为既定事实而忽略
  • 本文贡献: 提出"光速 ISP(Speed-of-Light ISP)"概念, 证明通过基础设施变革降低延迟是可行的

(2) 工业界基础设施项目

  • Google X (Project Taara): 基于自由空间光学(Free-space optics)技术, 在印度和非洲部署, 旨在通过"光束"提供快速, 可负担的互联网接入
  • Facebook Connectivity:
    • Magma / Rural Access: 针对农村地区, 提供软件, 硬件及商业模式框架
    • Terragraph: 利用毫米波(mmWave)短跳技术, 解决城市和郊区的"最后一公里"连接问题
  • 与 cISP 的关系: 这些项目证明了视距(Line-of-Sight)网络的可行性. cISP 设计具有灵活性, 可兼容光纤, 微波, 毫米波和自由空间光学等多种介质

(3) “新太空”卫星网络

  • 代表项目：Starlink, Kuiper, Telesat 等
  • 定位差异：
    • LEO：目前主要侧重于“最后一公里（Last-mile）”接入，而非长途骨干传输
    • cISP：专注于“长途（Long-haul）”低延迟传输
  • 延迟现状：Starlink 声称的最后一公里 RTT 为 31ms，这比之前的理论模拟值高出 3.8 倍，说明其目前尚未针对延迟进行极致优化

(4) Cost Comparison

  • 尽管 cISP 与 LEO 卫星的目标不同（长途 vs. 最后一公里），但在每比特成本上具有可比性:
    • Starlink：约 $0.36/GB（基于 $99/月费和平均户用流量估算）
    • cISP：约 $0.33/GB（若由 American Tower 等现有塔商部署）
  • 结论：cISP 在单位流量成本上与卫星网络相当，但在绝对总成本上低一个数量级（当然总带宽也相应较低）

(5) 现有的低延迟利基市场

  • 高频交易 (HFT)：金融市场的点对点微波链路
  • 海底光缆：旨在缩短特定路由的隔离光缆项目
  • 局限性：这些努力目前仅存在于特定的小众领域，未惠及广域互联网
类别 代表项目/技术 主要目标/特点 与 cISP 的关系/对比
学术研究 传输/路由协议优化 优化软件与协议栈，测量延迟 cISP 关注被学术界忽视的物理基础设施层延迟
自由空间/无线项目 Google Taara
FB Terragraph		 自由空间光学、毫米波；
侧重农村或城市“最后一公里”		 证明了 LoS 技术的可行性；
cISP 架构灵活，可融合这些技术
LEO 卫星网络 Starlink
Kuiper		 低轨卫星群；
侧重全球覆盖和接入		 Starlink 侧重接入，且延迟尚未优化；
cISP 侧重骨干长途，且单位成本更低 ($0.33 vs $0.36 /GB)
利基市场 HFT 微波链路
海底光缆		 极端低延迟；
服务于金融交易或特定路由		 cISP 将这种“特权”技术扩展到广域互联网，服务大众