没细看，忘记了论文内容CAPS: Coding-Based Adaptive Packet Spraying to Reduce Flow Completion Time in Data Center

1 CAPS: Coding-Based Adaptive Packet Spraying to Reduce Flow Completion Time in Data Center

CAPS：基于编码的自适应数据包喷洒，以减少数据中心的流完成时间

Abstract— Modern data-center applications generate a diverse
mix of short and long flows with different performance
requirements and weaknesses. The short flows are typically delay-
sensitive but to suffer the head-of-line blocking and out-of-order
problems. Recent solutions prioritize the short flows to meet their
latency requirements, while damaging the throughput-sensitive
long flows. To solve these problems, we design a Coding-based
Adaptive Packet Spraying (CAPS) that effectively mitigates the
negative impact of short and long flows on each other. To exploit
the availability of multiple paths and avoid the head-of-line
blocking, CAPS spreads the packets of short flows to all paths,
while the long flows are limited to a few paths with Equal
Cost Multi Path (ECMP). Meanwhile, to resolve the out-of-
order problem with low overhead, CAPS encodes the short flows
using forward error correction (FEC) technology and adjusts
the coding redundancy according to the blocking probability.
Moreover, since the coding efficiency decreases when the coding
unit is too small or large, we demonstrate how to obtain the
optimal size of coding unit. The coding layer is deployed between
the TCP and IP layers, without any modifications on the existing
TCP/IP protocols. The test results of NS2 simulation and small-
scale testbed experiments show that CAPS significantly reduces
the average flow completion time of short flows by∼30%-70%
over the state-of-the-art multipath transmission schemes and
achieves the high throughput for long flows with negligible traffic
overhead.
Index Terms— Data center, TCP, packet spray, multipath.

摘要-现代数据中心应用程序生成了各种长短流的混合，具有不同的性能要求和弱点。短流程通常对延迟敏感，但会遇到队头阻塞和无序问题。最近的解决方案优先考虑短流以满足其延迟要求，同时破坏吞吐量敏感的长流。为了解决这些问题，我们设计了一种基于编码的自适应包喷洒（CAPS），有效地缓解了短流和长流之间的负面影响。为了利用多条路径的可用性并避免线头阻塞，CAPS将短流的数据包传播到所有路径，而长流则被限制为具有相同成本的多路径（ECMP）的少数路径。同时，为了解决低开销的无序问题，CAPS采用前向纠错（FEC）技术对短流进行编码，并根据阻塞概率调整编码冗余。此外，由于当编码单元太小或太大时，编码效率会降低，因此我们演示了如何获得最佳的编码单元大小。编码层部署在TCP层和IP层之间，对现有TCP/IP协议没有任何修改。NS2模拟和小型试验台实验的测试结果表明，CAPS显著缩短了短流的平均流完成时间∼比最先进的多路径传输方案高出30%-70%，并在流量开销可忽略不计的情况下实现长流量的高吞吐量。

索引术语-数据中心、TCP、数据包喷洒、多路径。

短流对延迟敏感，有线路阻塞和无序问题。为了解决这些问题，我们设计了一种基于编码的自适应包喷洒（CAPS）。

1为了利用多条路径的可用性并避免线头阻塞，CAPS将短流的数据包传播到所有路径，而长流则被限制为具有相同成本的多路径（ECMP）的少数路径。

2同时，为了解决低开销的无序问题，CAPS采用前向纠错（FEC）技术对短流进行编码，并根据阻塞概率调整编码冗余。

3此外，由于当编码单元太小或太大时，编码效率会降低，因此我们演示了如何获得最佳的编码单元大小。编码层部署在TCP层和IP层之间，对现有TCP/IP协议没有任何修改。

实验表明，CAPS显著缩短了短流的平均流完成时间∼比最先进的多路径传输方案高出30%-70%，并在流量开销可忽略不计的情况下实现长流量的高吞吐量。

总：用CAPS技术解决了短流的线路阻塞和无序问题，实验表明比以前好。

I. INTRODUCTIONI
N MODERN data centers, a variety of cloud-based appli-
cations such as web search and social networking are
deployed by a large number of online service providers like
Google and Facebook. To obtain better user experience and
financial revenue, how to achieve the low latency and high
throughput becomes a crucially important issue.

简介在现代数据中心，大量在线服务提供商（如谷歌和Facebook）部署了各种基于云的应用程序，如网络搜索和社交网络。为了获得更好的用户体验和财政收入，如何实现低延迟和高吞吐量成为一个至关重要的问题。

Random Packet Spraying (RPS) [1] splits each flow into
packets and spreads the packets to all available paths to
achieve high efficiency and easy deployment in the multipath
topologies such as FatTree and Clos, which are widely used in
the large-scale data center networks (DCNs). In recent years,
the packet spraying technique has already been implemented
on the commodity switches (e.g. Cisco [2]).

随机数据包喷洒（RPS）[1]将每个流拆分为数据包，并将数据包散布到所有可用路径，以实现在大规模数据中心网络（DCN）中广泛使用的FatTree和Clos等多路径拓扑中的高效率和轻松部署。近年来，包喷洒技术已经在商品交换机（如Cisco[2]）上实现。

点评：讲RPS技术的好处；

However, RPS neglects the important traffic characteristic
that the short and long Transmission Control Protocol (TCP)
flows are mixed in DCNs. The data center traffic can be
characterized by heavy-tailed distribution [3], [4], that is,
around 90% of data is provided by only around 10% of TCP
flows, and about 90% of TCP flows provide only about 10%
of data. Furthermore, most short flows belong to the delay-
sensitive applications, while the majority of long flows are
throughput-sensitive. Since RPS does not distinguish between
short and long flows, it inevitably leads to the negative
interaction between the two kinds of flows.

然而，RPS忽略了DCN中长短传输控制协议（TCP）流混合的重要流量特性。数据中心流量的特征是重尾分布[3]，[4]，即，大约90%的数据仅由大约10%的TCP流提供，而大约90%的TCP流仅提供大约10%的数据。此外，大多数短流属于延迟敏感的应用程序，而大多数长流是吞吐量敏感的。由于RPS不区分短流和长流，因此不可避免地导致这两种流之间的负相互作用。

点评：RPS忽略了数据中心流量的特征：重尾分布；RPS不区分短流和长流；

The following two issues emerge when the packets of short
and long flows are mixed over the multiple paths by RPS
scheme. The first one is the head-of-line blocking. When
the switch buffer is occupied by the packets of long flows,
the short flows have to experience large queueing delay,
leading to the long-tailed flow completion time (FCT). Sec-
ondly, RPS randomly spreads the packets into different paths,
possibly resulting in the TCP out-of-order problem.

当短流和长流的数据包通过RPS方案在多条路径上混合时，会出现以下两个问题。第一个是线路阻塞的头部。当交换缓冲区被长流数据包占用时，短流必须经历较大的排队延迟，从而导致长尾流完成时间（FCT）。第二，RPS将数据包随机分散到不同的路径，可能导致TCP无序问题。

点评：RPS方案不好，短流和长流的数据包在多条路径上混合时，会出现两个问题。

Coding is a very powerful scheme to address these issues.
At the sender, the source packets are encoded and scattered to
multiple paths. Though some packets experience the head-of-
line blocking or out-of-order, once sufficient encoded packets
arrive at the receiver, the original packets can be recovered
immediately.In this paper, we propose a coding-basedadaptive
packet spraying (CAPS), which successfully integrates coding
into packet spraying and efficiently avoids head-of-line block-
ing and packet reordering. To improve the coding efficiency,
we provide quantitative analysis of the number of redundant
packets and the size of coding unit. Moreover, the transparent
coding layer only needs to be deployed between the TCP and
Internet Protocol (IP) layers at the end hosts, while making
no modifications on the existing TCP/IP protocols.
In summary, our major contributions are as follows:

编码是解决这些问题的一个非常强大的方案。在发送方，源数据包被编码并分散到多条路径。尽管有些数据包遇到了线路阻塞或无序，但一旦足够的编码数据包到达接收器，原始数据包可以立即恢复。在本文中，我们提出了一种基于编码的自适应包喷洒（CAPS），它成功地将编码集成到包喷洒中，并有效地避免了行首阻塞和包重排序。为了提高编码效率，我们提供了对冗余数据包数量和编码单元大小的定量分析。此外，明显的编码层只需要部署在终端主机的TCP和Internet协议（IP）层之间，而不需要修改现有的TCP/IP协议。总之，我们的主要贡献如下：

点评：基于编码的自适应包喷洒（CAPS）解决了行首阻塞和包重排序。避免了队头阻塞和包重排序。

We conduct an extensive simulation-based study to ana-
lyze two key issues with multipath transmission: the
short flows experience large FCT due to the head-of-line
blocking caused by the long flows, and the coexisting of
short and long flows leads to the reordering problem.

我们进行了广泛的基于仿真的研究，以分析多径传输的两个关键问题：短流由于长流导致的队头阻塞而经历较大的FCT，以及短流和长流共存导致重新排序问题。

点评：多径传输会导致两个问题，1队头阻塞 2重新排序

We propose a multipath transport scheme CAPS, which
randomly scatters the encoded packets of short flows to
all paths and transmits the long flows to a few paths.
Specifically, we design the coding layer, which ratio-
nally adjusts the number of redundant packets to resolve
the out-of-order and head-of-line blocking problems.
To improve the link utilization, CAPS swiftly scatters the
packets of long flows to the unused paths by short flows.

我们提出了一种多径传输方案CAPS，该方案将短流的编码包随机分散到所有路径，并将长流传输到少数路径。具体来说，我们设计了编码层，合理调整冗余数据包的数量，以解决无序和行首阻塞问题。为了提高链路利用率，CAP通过短流将长流的数据包迅速分散到未使用的路径。

点评：提出了多径传输方案CAPS，该方案将短流的编码包随机分散到所有路径，并将长流传输到少数路径。

To achieve high coding efficiency, we firstly demonstrate
experimentally and theoretically why the FCT perfor-
mance is impacted by the size of coding unit. Then
we give the mathematical analysis on how to obtain the
optimal size of coding unit and verify the accuracy of our
analysis results.

为了获得较高的编码效率，我们首先从实验和理论上论证了编码单元大小对FCT性能的影响。然后对如何获得最佳编码单元大小进行了数学分析，并验证了分析结果的准确性。

点评：找最佳的编码单元大小

By using both Network Simulator 2 (NS2) simulations
and small-scale testbed experiments, we demonstrate that
CAPS performs remarkably better than the state-of-the-
art multipath transmission schemes. Especially, CAPS
greatly reduces the average FCT (AFCT) of short flows
by∼30%-70% under high workload. Meanwhile, CAPS
improves throughput of long flows by up to∼45% and
∼35% over RepFlow [5] and Freeway [6], respectively.

通过使用网络模拟器2（NS2）仿真和小规模试验台实验，我们证明CAPS的性能明显优于现有的多径传输方案。特别是，CAPS通过以下方式大大降低了短流的平均FCT（AFCT）∼在高工作负荷下30%-70%。同时，CAPS将长流的吞吐量提高了多达∼45%及∼分别超过RepFlow[5]和高速公路[6]的35%。

点评：实验表明CAPS好；

The rest of the paper is organized as following. In Section II,
we present the related works. In Section III and IV,
we respectively describe our design motivation and overview.
In Section V and VI, we introduce packet spraying and
coding of CAPS, respectively. We discuss the implementa-
tion in Section VII. In Section VIII and IX, we show the
testbed experimental and NS2 simulation results, respectively.
We conclude the paper in Section X. The appendixes provide
supplementary experiments related to Incast scenarios and
priority queues.

论文的其余部分组织如下。在第二节中，我们介绍了相关的工作。在第三节和第四节中，我们分别描述了我们的设计动机和概述。在第五节和第六节中，我们分别介绍了CAP的包喷洒和编码。我们将在第七节中讨论实施。在第八节和第九节中，我们分别展示了试验台实验和NS2模拟结果。我们在第十节中总结了本文。附录提供了与Incast场景和优先级队列相关的补充实验。

APPENDIX appendix附录

X. CONCLUSION
To mitigate the negative impact of short and long flows
on each other in data center networks, we propose CAPS,
a coding-based adaptive packet spraying design that reduces
the flow completion time for short flows and guarantees the
throughputs for long flows. CAPS utilizes the FEC code to
encode only the short flows and spreads the packets of short
flows to all equal cost multipath. CAPS limits the long flows
when coexisting with the short flows to avoid the head-of-
line problem and scatters the packets of long flows to the
unused paths by the short flows to achieve high throughput.
We evaluate CAPS with both NS2 simulations and small-scale
Mininet testbed. The results indicate that CAPS significantly
reduces the AFCT by∼30%-70% for short flows and achieves
high throughput for long flows with negligible traffic overhead.

X.结论

为了缓解数据中心网络中短流和长流相互之间的负面影响，我们提出了CAPS，这是一种基于编码的自适应数据包喷洒设计，可减少短流的流完成时间，并保证长流的吞吐量。CAPS利用FEC代码仅对短流进行编码，并将短流分组传播到所有等成本多路径。CAPS当短流与长流共存时会限制长流，以避免队头阻塞问题，并通过短流将长流分组分散到未使用的路径，以实现高吞吐量。我们使用NS2模拟和小型Mininet试验台评估CAPS。结果表明，CAPS显著降低AFCT∼短流量为30%-70%，长流量为高吞吐量，流量开销可忽略不计。

点评：基于编码的自适应数据包喷洒CAPS仅对短流进行编码，不对长流编码，将短流分散到等价的多个路径，将长流分散到未使用的路径；这样可以避免队头阻塞问题和实现高吞吐量。

想法：

1 数据中心网络的流量是重尾分布，

2 我能不能对CAPS这个方法进行优化；提高短流的FCT或是长流的吞吐量；

3 我能不能换个方法，不用CAPS这个方法，换个其它的方法；