跳转至

INTRODUCTION

Content Distribution (or delivery) Networks (CDNs), which consist of a considerable number of geo-distributed cloud-based cache servers, aim at providing high network availability and low content access latency globally. The global CDN delivers a large fraction of total Internet traffic today [6]. Therefore, optimizing the network performance of CDNs can significantly improve the quality of experience (QoE) of a variety of applications built upon CDNs (e.g., Web services and Video-on-Demand (VoD)).

内容分发网络（CDN）由大量地理分布的基于云的缓存服务器组成，旨在提供高网络可用性和低内容访问延迟。如今，全球CDN传输了很大一部分互联网总流量。因此，优化CDN的网络性能可以显著提高构建在CDN之上的各种应用程序（例如，Web服务和视频点播（VoD））的体验质量（QoE）。

Content access latency, which is typically defined as the time consumption of delivering requested objects to end users, is one of the most important performance metrics in CDNs. In practice, the content access latency can refer to the page load time of a website, or the video initialization time for VoD, depending on the concrete application types. Critical to the low-latency story of CDNs is that users can access content replicas cached on geo-distributed servers close to end users. However, although CDNs hold great promise, our analysis on a large-scale CDN trace collected from seven major CDN operators across 183 countries reveals that: from a global perspective, even though today’s CDNs have been deployed for many years with a considerable amount of cache servers around the world, there are still a large portion of CDN users suffering from high round-trip time (RTT) even to their closest cache server, which can probably result in long content access latency. Our further analysis in §II identifies that such high access latency is more prevalent in remote or rural areas, due to two important reasons: (i) the insufficient deployment of cloud infrastructures, and (ii) meandering terrestrial routes from clients to cloud servers (e.g., prolonged Internet paths caused by remote peering [26]).

内容访问延迟通常定义为将请求的对象传递给最终用户所消耗的时间，是CDN中最重要的性能指标之一。在实践中，内容访问延迟可以指网站的页面加载时间，或者VoD的视频初始化时间，具体取决于具体的应用程序类型。 CDN低延迟的关键在于用户可以访问缓存在靠近最终用户的地理分布式服务器上的内容副本 。

然而，尽管CDN具有很大的前景，但我们对从183个国家/地区的七个主要CDN运营商收集的大规模CDN跟踪数据的分析表明：从全球角度来看，即使今天的CDN已经部署多年，并在世界各地拥有大量的缓存服务器，但仍然有很大一部分CDN用户即使到离他们最近的缓存服务器也遭受较高的往返时间（RTT），这可能会导致较长的内容访问延迟。

我们在第二节中的进一步分析表明，由于两个重要的原因，这种 高访问延迟在偏远或农村地区更为普遍：（i）云基础设施的部署不足，以及（ii）从客户端到云服务器的蜿蜒陆地路线（例如，远程对等互连导致的网络路径延长）。

However, none of the above root causes is easy to address in existing terrestrial networks. First, the deployment of today’s Internet is essentially an uneven network, where network resources are aggregated in many developed regions (e.g., “hot areas”). In remote and other under-developed areas, Internet access is limited. Provisioning and maintaining cloud servers in such areas to improve network performance are technically and economically challenging. Second, terrestrial Internet is divided into many autonomous systems (ASes). Due to the concrete routing policies of different independent ASes, inter-AS routes might be tortuous, resulting in meandering routes between users and assigned cloud servers, which further increase the client-to-cache latency.

然而，在现有的陆地网络中，上述任何一个根本原因都不容易解决。首先，当今互联网的部署本质上是一个不均衡的网络，网络资源集中在许多发达地区（例如，“热点地区”）。在偏远和其他欠发达地区，互联网接入受到限制。在这种地区配置和维护云服务器以提高网络性能在技术上和经济上都具有挑战性。其次，陆地互联网被划分为许多自治系统（AS）。由于不同独立AS的具体路由策略，AS间路由可能很复杂，导致用户和分配的云服务器之间的路由迂回，这进一步增加了客户端到缓存的延迟。

Emerging mega-constellations with thousands of satellites flying in low earth orbit (LEO) raise a new opportunity to optimize the network performance of CDNs globally. As today’s space-crafts have evolved rapidly in recent years, modern satellites will be equipped with high-throughput communication components [32], [36] and high-capacity storage [37], [38]. Such satellites have the potential to act as “cache in space”. Intuitively, caching content replicas on emerging LEO constellations close to users is a promising approach to enable low latency pervasively [22]. However, fully exploiting the potential of mega-constellations still needs to address several fundamental challenges. First, satellites are fundamentally mobile, and moving at a high velocity in their orbits. How should we properly select LEO satellites to cache content replicas and avoid the impact of intermittent connectivity? Second, it is more expensive to carry CDN traffic over satellites than that over terrestrial networks. How should we judiciously assign user requests to a satellite or a cloud cache server in a cost-effective manner?

新兴的拥有数千颗在低地球轨道（LEO）运行的卫星的巨型星座为优化全球CDN的网络性能提供了一个新的机会。随着当今的航天器近年来发展迅速，现代卫星将配备高吞吐量通信组件，和高容量存储，。这些卫星有可能充当“太空缓存”。直观地，在靠近用户的新兴LEO星座上缓存内容副本是一种很有前途的实现普遍低延迟的方法。

然而，充分利用大型星座的潜力仍然需要解决几个根本性的挑战。首先， 卫星本质上是移动的，并且以高速在其轨道上移动。我们应该如何正确选择LEO卫星来缓存内容副本 ，并避免间歇性连接的影响？其次， 通过卫星传输CDN流量比通过地面网络传输更昂贵 。我们应该如何以具有成本效益的方式将用户请求明智地分配给卫星或云缓存服务器？

We propose STAR FRONT, a content distribution framework that cooperatively leverages cache servers in both LEO satellites and terrestrial clouds to optimize the content access latency in a cost-effective manner. In particular, STAR FRONT adopts multiple techniques to accomplish the goal of cost-effective wide-area content distribution. First, STAR FRONT constructs a dynamic satellite-cloud model which captures the time-varying accessibility and network performance of the satellite-cloud integrated architecture. The model is built based on the information of cloud distribution, predictable satellite trajectory together with their estimated performance and pricing policies specified by satellite or cloud operators.

我们提出了STAR FRONT，这是一个内容分发框架，它协同利用LEO卫星和陆地云中的缓存服务器，以经济高效的方式优化内容访问延迟。特别是，STAR FRONT采用多种技术来实现经济高效的广域内容分发目标。

首先，STAR FRONT构建了一个动态 卫星-云 模型，该模型捕获了 卫星-云 集成架构的随时间变化的可访问性和网络性能。该模型是基于云分布信息、可预测的卫星轨迹以及卫星或云运营商指定的估计性能和定价策略而构建的。

Second, based on the integrated satellite-cloud architecture, the STAR FRONT framework enables three forms of request assignments for users in different regions. Specifically, a user request can be: (i) directly assigned to a cloud server (i.e., a terrestrial cache) via terrestrial networks like that in existing cloud-based CDNs; (ii) assigned to a cloud server through lowlatency space paths constructed by a sequence of satellites; or (iii) assigned to a satellite cache if the nearest cloud is still too far away. The above approaches of request assignment involve different latency performance and corresponding storage and traffic costs in practice.

其次，基于集成的卫星-云架构，STAR FRONT框架为不同地区的用户实现了三种形式的请求分配。具体来说，用户请求可以：

（i）像现有的基于云的CDN一样，通过地面网络直接分配给云服务器（即地面缓存）；

（ii）通过由一系列卫星构建的低延迟空间路径分配给云服务器；

（iii）如果最近的云仍然太远，则分配给卫星缓存。实际上，上述请求分配方法涉及不同的延迟性能以及相应的存储和流量成本。

Third, STAR FRONT incorporates a collection of content placement and request assignment algorithms to achieve cost-effective content distribution. Specifically, STAR FRONT judiciously pushes contents and places replicas on available cloud or satellite servers, and assigns user requests to proper cache servers to meet the latency requirement of different applications, while minimizing the total cost of content distribution via satellite-cloud cooperation.

第三，STAR FRONT结合了一系列内容放置和请求分配算法，以实现经济高效的内容分发。具体来说，STAR FRONT明智地推送内容并将副本放置在可用的云或卫星服务器上，并将用户请求分配给适当的缓存服务器，以满足不同应用程序的延迟要求，同时最大限度地降低通过 卫星-云 合作进行内容分发的总成本。

We build a testbed to simulate a large number of cloud caches and dynamic LEO satellite caches to evaluate the proposed framework. We also implement a STAR FRONT prototype based on Apache Traffic Server (ATS) [4] running upon the testbed. Further, we conduct trace-driven evaluations based on three state-of-the-art constellations covering a collection of geo-distributed vantage points to evaluate the effectiveness of STAR FRONT. Our evaluation results reveal that by integrating satellites and clouds for content distribution globally, STAR FRONT can satisfy various latency requirements of applications, and outperform existing cloud-only approaches in terms of content access latency with acceptable operational cost and resilience to intermittent connectivity. Moreover, we find that the constellation structure can significantly affect the achievable performance of STAR FRONT. In particular, constellations with lower altitude and equipped with intersatellite links (ISLs) can further reduce more content access latency for terrestrial users, since lower altitude indicates lower propagation delay, and high-speed ISLs enable near-optimal space paths for users to fetch contents from cloud servers over LEO satellites.

我们构建了一个测试平台来模拟大量的云缓存和动态LEO卫星缓存，以评估所提出的框架。我们还基于在测试平台上运行的Apache Traffic Server（ATS）实现了STAR FRONT原型。此外，我们基于涵盖一系列地理分布有利位置的三种最先进的星座进行了跟踪驱动的评估，以评估STAR FRONT的有效性。我们的评估结果表明，通过集成卫星和云进行全球内容分发，STAR FRONT可以满足各种应用程序的延迟要求，并且在内容访问延迟方面优于现有的纯云方法，同时具有可接受的运营成本和对间歇性连接的弹性。此外，我们发现星座结构会显著影响STAR FRONT的可实现性能。特别是，具有较低高度并配备星间链路（ISL）的星座可以进一步减少地面用户的更多内容访问延迟，因为较低的高度表示较低的传播延迟，而高速ISL使用户能够通过LEO卫星从云服务器获取内容的近乎最佳的空间路径。

Taken together, this paper makes three key contributions:

• (i) We identify and analyze the high access latency problem in existing CDNs through a measurement study on seven commercial CDN operators, and expose the feasibility as well as the challenges of exploiting emerging LEO mega-constellations to assist pervasive and lowlatency content distribution globally (§II).

• (ii) We formulate the cost-effective content distribution (CECD) problem in the satellite-cloud integrated environment, and present STAR FRONT, a content distribution framework that cooperatively leverages the storage and network capabilities in both clouds and LEO satellites to judiciously optimize the content access latency on a global scale and in a cost-effective manner (§III,§IV,§V).

• (iii) We implement a prototype of STAR FRONT , and evaluate the effectiveness of STAR FRONT on satisfying various latency requirements for geo-distributed users via extensive trace-driven simulations upon the prototype (§VI). The CDN trace data used in our evaluation is now available at: https://github.com/SpaceNetLab/STAR FRONT.

总而言之，本文做出了三个主要贡献：

  • (i) 我们通过对七个商业CDN运营商的测量研究，识别并分析了现有CDN中的高访问延迟问题，并揭示了利用新兴的LEO巨型星座来协助全球普遍和低延迟内容分发的可行性以及挑战（第II节）。

  • (ii) 我们在卫星-云集成环境中制定了具有成本效益的内容分发（CECD）问题，并提出了STAR FRONT，这是一个内容分发框架，它协同利用云和LEO卫星中的存储和网络功能，以明智地优化全球范围内的内容访问延迟，并以具有成本效益的方式（第III，第IV，第V节）。

  • (iii) 我们实现了STAR FRONT的原型，并通过在原型上进行的大量跟踪驱动模拟，评估了STAR FRONT在满足地理分布式用户的各种延迟要求方面的有效性（第VI节）。我们在评估中使用的CDN跟踪数据现在可在以下网址获得：https://github.com/SpaceNetLab/STARFRONT。