云计算部署探析：Amazon EC2与Azure
需积分: 10 116 浏览量
更新于2024-09-10
收藏 1.72MB PDF 举报
"Amazon EC2 和 Azure 是两个领先的云计算平台，被广泛用于托管互联网服务。这篇论文深入探讨了在 EC2 和 Azure 上部署 Web 服务的情况，揭示了现代 Web 服务如何利用云技术，并提出了改进部署策略的可能性。研究发现，有4%的 Alexa 百万顶级网站使用 EC2 或 Azure，同时存在一些常见的云服务前端部署模式。通过更好地利用 EC2 的区域多样性，服务可以显著提高其广域网性能和容错能力。这些分析基于多个新的数据集，包括超过3400万个 Alexa 网站的 DNS 记录以及一个大型大学网络的包捕获数据。"
本文首先强调了 Amazon EC2 和 Windows Azure 在云计算领域的主导地位，指出越来越多的互联网服务选择在这些平台上进行托管。由于缺乏对云服务使用情况的深入研究，论文旨在填补这一空白。
研究方法主要包括详尽的测量分析，研究人员收集并分析了大量的数据，包括对 Alexa 排名前百万的网站中的 EC2 和 Azure 使用情况进行统计，以及对一个大型学术网络的网络流量进行包捕获，以便了解实际的流量模式和性能表现。
结果表明，尽管 EC2 和 Azure 的使用比例相对较低（4%），但它们已经成为许多顶级网站不可或缺的一部分。此外，研究揭示了几种常见的云服务前端部署策略，这些策略可能反映了业界的最佳实践或者是为了优化特定的性能指标。
论文特别关注了如何利用 EC2 的地理分布来提高服务的可用性和性能。研究发现，通过更好地利用不同地区的服务器，服务提供商可以显著减少延迟，增强故障恢复能力。这表明云服务用户应该更加重视跨地域的负载均衡和冗余策略，以提升用户体验和系统稳定性。
最后，论文提供的新数据集对于未来的研究工作极具价值，它们为理解云服务部署的复杂性提供了基础，有助于研究人员和从业者更深入地探索云环境下的性能优化、安全性和其他关键问题。这篇论文不仅提供了关于 EC2 和 Azure 使用现状的宝贵见解，还为改善云服务部署提供了指导性的建议。
Figure 2: PlanetLab nodes used for DNS lookups
nodes, which are spread across North America, South America,
Europe, Asia, and Australia. The queries were performed March
27-29, 2013. These distributed DNS queries help ensure that we
gather a comprehensive set of DNS records for each cloud-using
subdomain and capture any geo-location-specific cloud usage.
We refer to the list of cloud-using subdomains, and their associ-
ated DNS records, as the Alexa subdomains dataset.
Packet Capture Dataset. Our second primary dataset is a se-
ries of packet traces captured at the border of the Universityof
Wisconsin-Madison campus network
3
.WecapturedfullIPpackets
whose source or destination IP address fell wi t hin the publicad-
dress ranges published by EC2 and Azure. The capture was per-
formed from Tuesday, June 26 to Monday, July 2, 2012 giving
us a full week of traffic and a total of 1.4TB of data. The total
Internet traffic averaged approximately 7Gbps during the capture,
with about 1% of the t raffic going to/coming from EC2 or Azure.
Due to the relatively low rate of traffic being captured, no loss oc-
curred during the capture process (according to tcpdump and coun-
ters reported by the border router). To protect user privacy,we
anonymized the IP addresses of clients within the universitynet-
work, and we only report aggregate statistics.
Since our traces contain full packets, we were able to performan
in-depth analysis of network and transport layer information (e.g.,
IP addresses, protocols, ports), application layer information (e.g.,
HTTP hostnames, HTTP content-type, HTTPS certificates), and
packet payloads. We extracted relevant information from thetraces
using Bro [33], a network monitoring and traffic analysis tool. We
refer to these traces as the packet capture dataset.
We recognize that a packet trace from a single campus vantage
point may not reflect the “typically” usage patterns of services de-
ployed in IaaS clouds. Correspondingly, we only leverage thepack-
et capture for analysis which cannot be conducted using our Alexa
subdomains dataset—namely, protocol usage (§3.1), popularity es-
timates based on traffic volume and flow counts (§3.2), and flow
characteristics (§3.3).
3. WEB-FACING CLOUD TENANTS
In this section, we explore what applications are being hosted
on public IaaS clouds. We start by analyzing the packet capture
to identify the types of applications being hosted. This analysis
suggests (unsurprisingly) that web applications representalarge,
important set of cloud tenants. We then turn to examining which
of the most popular websites are using clouds. We view popularity
both globally, via the Alexa top website rankings, and locally, via
the volume of traffic associated with each domain in the packetcap-
ture. We also analyze the traffic patterns of cloud-using services,
including flow characteristics and content types served.
3
The university has seven /24 IP blocks and one /16 IP block
Cloud Bytes Flows
EC2 81.73 80.70
Azure 18.27 19.30
Total 100 100
Table 1: Percent of traffic volume and percent of flows associated
with each cloud in the packet capture.
EC2 Azure Overall
Protocol Bytes Flows Bytes Flows Bytes Flows
ICMP 0.01 0.03 0.01 0.18 0.01 0.06
HTTP (TCP) 16.26 70.45 59.97 65.41 24.24 69.48
HTTPS (TCP) 80.90 6.52 37.20 6.92 72.94 6.60
DNS (UDP) 0.11 10.33 0.10 11.59 0.11 10.58
Other (TCP) 2.40 0.40 2.41 1.10 2.40 0.60
Other (UDP) 0.28 0.19 0.31 14.77 0.28 3.00
Total 100 100 100 100 100 100
Table 2: Percent of traffic volume and percent of flows associated
with each protocol in the packet capture.
3.1 Protocols and Services
We first examine the fraction of bytes and flows in the packet
capture that are associated with each cloud (Table 1). We only
consider flows that were initiated within the university and destined
for EC2 or Azure. We observe that the majority of cloud traffic,
both as measured by volume and number of flows, is EC2-related:
81.73% of bytes (80.70% of flows) are associated with EC2, while
Azure accounts for 18.27% of bytes (19.30% of flows).
Next, we use the packet capture to study the application-layer
protocols used by cloud tenants. Table 2 shows the percentageof
bytes (and flows) using a specific protocol relative to the total num-
ber of bytes (and flows) for EC2, Azure, and the capture as a whole.
We observe that more than 99% of bytes in the packet capture are
sent and received using TCP, with less than 1% of bytes associated
with UDP or ICMP. The vast majority of this TCP t r affic is HTTP
and HTTPS. The proportion of HTTPS traffic is far higher than
that seen for general web services in the past (roughly 6% [18]);
as we will show later, HTTPS traffic is dominated by cloud stor-
age services. Interestingly, the majority of Azure’s TCP traffic is
HTTP (59.97%) while the majority of EC2’s TCP traffic is HTTPS
(80.90%)
The breakdown by flow count is less skewed towards TCP, with
UDP flows accounting for 14% of flows in the packet capture. This
is largely due to DNS queries, which account for 11% of flows but
carry few bytes.
As one would expect, public IaaS clouds are also used for non-
web-based services. In t he packet capture, we find a small fraction
of non-HTTP(S) TCP traffic and non-DNS UDP traffic going to
both EC2 and Azure. This traffic includes SMTP, FTP, IPv6-in-
IPv4, SSH, IRC, and other traffic that Bro could not classify.
Summary and implications. While we analyze a single vantage
point, our measurements suggest that web services using HTTP(S)
represent an important set of WAN-intensive cloud tenants. The
extent to which compute-intensive workloads (that may not result
in a large impact on network traffic) are prevalent as cloud tenants
remains an interesting open question. In the following sections we
dig into what tenants are hosting web services on public clouds as
well as diving deeper into their traffic patterns.
3.2 Popular Cloud-Using (Sub)Domains
Cloud-using Alexa domains. We now consider what subset of the
Alexa top 1 million websites use the cloud to (partly) host their
services. Recall that Alexa provides an estimate of the most pop-
剩余13页未读，继续阅读
2024-06-16 上传
141 浏览量
291 浏览量
点击了解资源详情
点击了解资源详情
点击了解资源详情
2021-05-10 上传
2012-10-26 上传
2012-04-01 上传
rongbear
- 粉丝: 0
我的内容管理
展开
最新资源
- 数学画图教具设计文档及应用指南
- SSH与WebService整合环境配置详解
- Java线程池基础教程与实例解析
- Notepad++ 2018及老版本编译工具链完整分享
- MFC实现圆弧扫描转换的图像处理技术
- 北大Hadoop环境下的数据库多表查询设计
- PHP表格讲习班：搜索栏导航与页面重定向
- 心理学教学辅助多媒体装置设计文档
- 三国游戏自动化工具开发：易语言实战攻略
- 深入解析Foxit PDF编辑器的强大功能
- C++仿FGO战斗逻辑的实现与代码分析
- React 练习项目深入探索
- MyEclipse10完整指南：构建WebService服务端和客户端
- Tensorflow.js实现的电晕面罩检测技术
- Spring Boot工具安装使用教程
- 圆木结构设计文档：木屋墙体的应用方案
