自适应与效用驱动的群通信服务扩展策略

153 浏览量更新于2024-07-15 收藏 1.2MB PDF 举报

本文档标题为"Scaling Group Communication Services with Self-adaptive and Utility-driven Message Routing"，是一篇研究论文，发表于MobileNetwAppl期刊，DOI为10.1007/s11036-011-0348-9。该研究主要关注在互联网边缘如何有效地处理大量组通信数据流，这是当前面临的一个挑战，因为这会给网络链接、资源需求以及路由节点能力带来巨大压力。传统上，许多研究集中在地理距离（Geo-distance）为基础的路由算法优化上，旨在减少路径长度或通过利用网络局部性来提升效率。然而，这些方法在应对网络动态变化时往往表现不足，尤其是在资源利用率和组通信效率方面存在局限。作者指出，现有的基于地理距离的路由协议在很多多播Overlay网络中并不高效。为解决这个问题，研究者提出了一种自适应和以效用为导向的消息路由策略。这种方法不再单纯依赖地理距离，而是转向一种更加灵活和智能化的方式，旨在优化资源分配，提升通信效率，并能动态适应网络条件的变化。效用驱动的路由机制可能包括但不限于流量负载均衡、带宽管理、冗余路径选择等，以确保在满足服务质量的同时，最小化网络资源的消耗。论文的核心贡献可能包括： 1. **效用模型的构建**：设计了一个新的效用函数，用于衡量不同路由决策的性能，结合了延迟、带宽、网络拥塞等因素，以实现更科学的路由决策。 2. **自适应路由算法**：开发了一种能根据网络状况实时调整的路由算法，它可以根据网络动态变化和效用模型动态选择最佳路径，从而提高整体系统性能。 3. **性能评估与优化**：论文可能详细分析了新方法在实际网络环境中的性能对比实验，展示了其在大规模组通信服务中的优越性，如降低延迟、减少拥塞、提升吞吐量等方面。 4. **潜在的应用场景**：讨论了这种自适应和效用驱动的路由策略在移动互联网、物联网、云计算等领域的潜在应用价值，以及对现有路由技术的改进和扩展。这篇论文提供了针对大规模组通信服务的一种创新解决方案，通过引入效用驱动和自适应策略，有效解决了传统路由方法在面对网络动态性和资源限制时的不足，为网络服务的可扩展性和效率提升开辟了新的研究方向。

Mobile Netw Appl

become shortcut nodes when they are not adjacent to

the given node.

Each node in the basic GeoCast maintains a list

peernodelist. Such list contains two kinds of nodes:

immediate neighbor node and shortcut node. Instead

of removing old immediate neighbors from the lists,

GeoCast keeps those nodes in peernodelists and uses

them to speed up the procedures of message deliv-

ery. For a node E

,itspeernodelist is denoted by

peernodelist(i) ={S

, S

,...,S

},whereQ is

defined by log

(G.w ∗G.h/(E

.R.w ∗E

.R.h)). For each

subset S

(1 ≤ j ≤ Q), it contains the nodes in the

geographical enclosing zone EZ

with the size of 1/2

of the geographical plane G. EZ

is defined by: EZ

x, y,w,h >,where(x, y) represents the coordinates of

top left vertex of enclosing zone and (w, h) refers to

the width and height of EZ

. Nodes in the subset S

are viewed as representatives of the enclosing zone

. If a message needs to be routed from node i

to a destination node contained in the enclosing zone

, it is likely that a shortcut node in the subset S

chosen as the next message forwarding hop.

Figure 1a provides a snapshot of GeoCast over-

lay network with three end system nodes. At this

point, there is no shortcut node included in the peern-

odelists of the nodes. For E

, peernodelist(2) ={S

{{E

, E

}}, where node E

and node E

are its neigh-

bors included in EZ

that is identical to the entire plane

G. With the arrival of node E

as shown in Fig. 1b,

node E

is no longer a neighbor of node E

and is

now recorded as a shortcut node in peernodelist(2).For

node E

, shortcuts are the nodes whose regions have

no intersection with region 2. Given the locations of

nodes, S

is represented by: S

={E

, E

},where

, E

∈ EZ

. Similarly, node E

becomes another

shortcut node of node E

after node E

joins the

network as shown in Fig. 1c. Now peernodelist(2) =

{{E

, E

}, {E

}},whereE

∈ EZ

and E

/∈ EZ

In this way, each node keeps building up its peernodelist

as the topology of network evolves with the arrival or

departure of nodes.

Different nodes may have their peernodelist in

different sizes. The exact length of the list for a node

is decided by the relative size of the region R

owned by E

. If the region R is 1/2

of the size

of the geographical plane, the list length is Q.This

allows to cover the entire geographical plane by the

shortcuts of E

according to the following equation:



i=1

1/2

+1/2

= 1 (The full version of the proof can

be found in our technical report [28]).

Let Gdist

i→j



− x

)

+(y

− y

)

be the geo-

distance between two nodes E

and E

on the space

G, where (x

, y

) and (x

, y

) are the unique identifier

of E

and E

respectively. Let destination be a point or

aregioninthespaceG[13]. In GeoCast, we consider

the destination as a spatial point in the plane and tag

its coordinates with a request message, consisting of

a spatial query point D(x, y), a description of request

data, information of the node who issued such request.

When the end system node that covers D(x, y) receives

this request, it sends the requested information related

to its region back to the request sender. In location-

based service (LBS), an example of such request is

“send me the current traffic state of I-85 on exit 94.”

When a node E

wants to route a message to the

node with the given destination coordinates, it first

checks if the coordinates are contained by the region it

owns. If not, it looks up the routing nodes in its peern-

odelist and chooses the node with the smallest Gdist

to the destination as its next routing hop to route the

message. This procedure is executed repeatedly until

the message reaches the node whose region covers the

destination. In SGD, the use of shortcuts ensures that

at each hop, the routing search space can be reduced

Fig. 1 A GeoCast overlay network

剩余20页未读，继续阅读

weixin_38519763

粉丝: 5
资源: 922

自适应与效用驱动的群通信服务扩展策略

Hands-On Machine Learning with Scikit-Learn and TensorFlow: Concepts, Tools, and

Hands-On Machine Learning with Scikit-Learn and TensorFlow [EPUB]

api-ms-win-shcore-scaling-l1-1-1.wim

api-ms-win-shcore-scaling-l1-1-1.dll

transient response

'train': ( "{cmd_mpi:s} nnp-scaling 100 > nnp-scaling-stdout.log 2> nnp-scaling-stdout.err; " "{cmd_mpi:s} nnp-train > nnp-train-stdout.log 2> nnp-train-stdout.err"), 'predict': '{cmd_mpi:s} nnp-dataset 0 > nnp-dataset-stdout.log 2> nnp-dataset-stdout.err'

N-dependent scaling of grand-averaged estimates是什么意思

为什么对ResNet34模型单独使用Linear scaling learning rate和共同使用Large-batch training与 Linear scaling learning rate和Large-batch training与No bias decay时模型性能反而下降

k8s operator

最新资源