优化能源效率：数据中心网络的吞吐量保障电源感知路由

下载需积分: 5 | PDF格式 | 1.55MB | 更新于2024-07-09 | 151 浏览量 | 举报

"通过吞吐量保证的电源感知路由来绿化数据中心网络" 这篇研究论文探讨了如何利用"通过吞吐量保证的电源感知路由"技术来优化数据中心网络的能效，从而实现“绿化”数据中心的目标。数据中心是现代互联网服务的基础，但它们的能源消耗巨大，对环境造成了一定的影响。传统的树形架构数据center网络在处理高流量和提供高可靠性服务时，往往存在能源效率低下的问题。论文中提到的Fat-Tree和BCube是近年来提出的新一代数据中心网络架构，它们通过过度设计（aggressively over-provisioning）来确保性能和可靠性。然而，这种方法虽然有效，但可能导致不必要的能源浪费。因此，研究者们转向了电源感知路由（Power-Aware Routing）技术，旨在通过智能地调整数据包的路由路径，减少能源消耗，同时保证网络的吞吐量，即数据传输速率。电源感知路由策略的核心在于理解网络设备的能耗模式，并据此制定路由决策。例如，当某些路径上的设备功耗较低时，路由协议可以优先选择这些路径，以减少整体能耗。同时，通过确保吞吐量保证，这种策略可以在不影响服务质量的前提下实现节能目标。论文还可能涉及网络功率消耗模型的研究，这是一个重要的理论基础，用于预测和分析不同路由策略对能源消耗的影响。通过建立准确的模型，研究人员可以评估和比较各种电源感知路由算法的性能，并找到最佳实践。在实际应用中，电源感知路由需要与数据中心的其他节能措施相结合，如冷热通道隔离、动态电压和频率调整（DVFS）、以及更高效的硬件设计。此外，还需要考虑网络动态性，因为流量模式会随时间变化，路由策略需要能够适应这些变化，以保持其能源效率。这篇研究论文探讨了电源感知路由在数据中心网络中的潜力，为降低能源消耗、提高绿色运营提供了新的视角。通过深入理解和实施这种技术，数据中心可以实现更高的能效比，为可持续的云计算服务打下坚实的基础。

展开

2.2. Throughput-guaranteed power-aware routing model

The objective of throughput-guaranteed power-aware

routing is to compute the transmission paths based on a gi-

ven trafﬁc matrix, so that the total power consumption of

switches involved in the routing paths is as little as possi-

ble while meeting predeﬁned network performance

thresholds. Throughout this paper we use network

throughput as a key performance metric, since it is the

most important metric for data-intensive computations.

For the convenience of presentation, we call our through-

put-guaranteed Power-aware Routing Problem PRP.

PRP: Assume there is a quadruplet of input parameters

(G,T, K, RP). G denotes the topology of a data center net-

work, which consists of servers and switches along with

their connections and link capacity. T is the network trafﬁc

matrix, which denotes the communication relationship be-

tween each pair of servers in the topology G. K and RP de-

note the predeﬁned thresholds of total network

throughput and network reliability respectively. The objec-

tive of PRP is to ﬁnd a routing R

⁄

for T, under the con-

straints of Eqs. (5)–(7). R denotes one of routing

selections from a candidate set R

which includes all pos-

sible routing paths for T, L(R) and X(R) denote the number

of switches and ports involved in R respectively, M(R) de-

notes the total network throughput of all ﬂows in T under

R, and Z(R) denotes the minimum number of available

paths for each ﬂow in T under R. We assume the ﬁxed

power consumption F of each switch is the same and all

switch ports in data center networks have the same capac-

ity and power consumption A, as current advanced data

center topologies, such as Fat-Tree and BCube, usually

use homogeneous commodity switches to interconnect

servers. We summarize the main notations used for our

power-aware routing model and algorithm in the Table 1.



¼ arg min

ðF  LðRÞþA  XðRÞÞ ð5Þ

MðRÞ P K; R 2R

ð6Þ

ZðRÞ P RP; R 2R

ð7Þ

The purpose of our model design is to ﬁnd the optimal

power-saving routing, in which the total power consump-

tion of switches involved is minimal while satisfying the

network throughput and reliability thresholds. In this pa-

per we focus on bandwidth-hungry batch-processing tasks

in data centers, such as MapReduce [1] and GFS [2]. For this

kind of application, total network throughput is a better

metric to reﬂect the computation progress, instead of the

constraints of single ﬂows. Therefore, we use the threshold

K to restrict the lower bound of the total throughput of all

ﬂows in the Eq. (6).

Also, it is necessary to ensure any network ﬂow identiﬁed

in the trafﬁc matrix will not be interrupted when using the

power-aware routing. As PRP is an NP-hard problem, it is dif-

ﬁcult to ﬁnd the optimal solution of the problem in polyno-

mial time, especially when the scale of the data center

network is large. We have proved NP-hardness of the prob-

lem by reducing the 0–1 Knapsack problem [10] to it. The de-

tails of NP-hardness proof can be found in the appendix.

It is worth noting that the model we use is different

from some previous similar work. In [7] and [41], each net-

work ﬂow has a ﬁxed trafﬁc rate, which is taken as the in-

put of the power minimization problem. In our PRP model,

the transfer rate of each ﬂow in the trafﬁc matrix is im-

pacted by the bandwidth competition among different

ﬂows, as the network is currently the bottleneck for dis-

tributed computation in data centers. Moreover, TCP trafﬁc

is dominant in current data center networks. TCP uses the

sliding window scheme and AIMD (Additive Increase Mul-

tiplicative Decrease) algorithm to control the transfer rate

of ﬂows. The work in [11] reveals that the rate of compet-

ing TCP ﬂows mainly relies on their RTT (Round-Trip Time).

In high-bandwidth and low-latency data center networks,

TCP ﬂows have a similar RTT value and share fairly the

available bandwidth of the bottleneck link. Therefore,

TCP follows the Max–Min fairness model [12] to assign

the transfer rate for each ﬂow in DCNs.

The throughput-guaranteed power-aware routing mod-

el is applied in data center networks, which have their own

characteristics compared with the Internet. First, current

advanced DCN architectures use an excessive number of

network resources to provide routing services and enjoy

abundant routing paths between each pair of servers. Also,

existing DCN architecture designs usually employ symmet-

ric topology structures [52] and homogeneous network de-

vices. Second, data center networks have more ﬂexible

trafﬁc control and management schemes, such as Open-

Flow [7], which make the design and implementation of

power-aware routing more tractable. Third, there are some

typical trafﬁc patterns in data centers, such as One-to-One,

One-to-Many and All-to-All trafﬁc patterns.

2.3. Related work

In the research area of green networking, some survey

papers [22,36] have summarized many novel network

power saving technologies appearing in recent years. In

this subsection, we only discuss typical data center net-

work architectures and representative power conservation

schemes used in Internet and data centers.

2.3.1. Data center network

Due to the well-known problem of the current practice

of tree topology, recently there are a bunch of proposals on

Table 1

Summary of main notations.

Notation Description

G topology of a data center network

T network trafﬁc matrix

K network throughput threshold

set of all possible routing paths

R a possible routing selection from R

⁄

optimal power-aware routing

F ﬁxed power consumption of switches

A power consumption of each switch port

L(R), X(R) number of switches and ports involved in R

M(R) total network throughput of all ﬂows under R

PR network performance threshold percentage

RP reliability requirement parameter

rate of ﬂow i

C capacity of a link

M. Xu et al. / Computer Networks 57 (2013) 2880–2899

2883

下载后可阅读完整内容，剩余19页未读，立即下载

身份认证购VIP最低享 7 折!

30元优惠券

weixin_38581405

粉丝: 2

优化能源效率：数据中心网络的吞吐量保障电源感知路由

提升多跳无线网络吞吐量：空间重用感知路由策略

云计算数据中心网络电源效率研究

网络编码支持的无线路由协议提升AdHoc网络吞吐量

CPTT：一种用于多跳无线网络的高吞吐量编码感知路由度量

车联网下基于网络编码的高吞吐量多径路由算法.pdf

无线传感器网络的流量整形网络编码感知路由

认知无线电自组织网络中具有稳定性约束的最大吞吐量路由

分布式网络编码感知路由CADR：提升无线网络性能的关键策略

机会路由EXOR：利用概率广播提升无线网络吞吐量

网络编码优化的无线Mesh路由协议：吞吐量提升20%

最新资源