The structure design of cloud computing data center network is the important
factor that must be considered. It can guarantee the high extensible ability
r and high utilization of resources in the cloud computing data center (DeCandia
et al., 2007; Irwin et al., 2004).
This study is aim for the extensibility in the structure of cloud computing
data center and saving needs of the energy and the new data center should have
the new characteristics, so it adopt already well-known data center structure,
on the basis of the Koch curve, proposed the snowflake structure which is a
new data center network structure, In the structure we take the scalability
of data centers into consideration, under the premise of lower quantity proportion;
it can realize the routing mechanism in a relatively short average path between
nodes and also with a smaller network expense (Dogan and
Osgunger, 2002; Erickson et al., 2009; Pinheiro
et al., 2007). The snowflake structure of cloud computing data center
network constructing method and its properties are discussed in the study, the
structure design between nodes in the routing protocols, algorithms and experiment
simulation are performed in order to verify the performance of the new structure
(Foster et al., 2008; Germain-Renaud
and Rana, 2009).
SEVERAL TYPICAL STRUCTURES OF DATA CENTER NETWORK
There are representative of the data center network architecture, such as a
Fat-Tree DCell, the BCube and the VL2 etc., it will analyze these structure
and their respective characteristics (Gilbert and Lynch,
The server can be divided into k groups of Fat-Tree structure, each subgroup
contains two layers of port number for the k/2 k switches, k/2 of the switch
port is connected to the k/2 hosts, respectively and each layer and switches
are connected; Core layer (k/2) 2 core switch port number is k, they most can
support the k3/4 hosts as shown in Fig. 1. Fat-tree abandoned
traditional data center and adopt the dedicated switch mode, switch to commercial
Ethernet switches, improves the ratio of larger. It can contain tens of thousands
of servers in the data center and provides high aggregate bandwidth, don't need
to modify the host network interface, operating system, you can build and compatible
with Ethernet, TCP/IP communication protocols as well.
|| Fat-tree structure in the cloud computing system
|| DCell_structure of the cloud computing system
Fat-link is equal to the number of tree layers, which makes maximum flow rate
equal to the maximum throughout of the core layer and it adopts the two levels
of routing table and the certain link error detection mechanism can realize
fault tolerant routing.
Removes the fat-tree structure tree structure upper link the limitation on
the throughput and can offer multiple parallel links for internal communication
between the nodes. But fat-tree extensibility is limited to the core switch
port number, which are frequently used at present is 48 core switch port with
10G capacity, in the three layers of tree structure, it can support 27648 hosts.
In the long run, the scale of data center is unable to meet the application
requirements, so the fat-tree exists defects with the lack of extensibility.
Fat-tree another disadvantage is the difference in fault tolerance, the manifestation
of lacking the ability to handle switch fault and fault tolerant routing protocol
is not strong. Research of MSRA show that fat-tree was highly sensitive to the
low level switch failures and it can seriously affects the performance of the
system. Because fat-tree is still the tree structure, so it has the defects
of tree structure essentially.
D-cell is a recursive definition of network structure, the use is located in
the i-1 layer D-cell and construct the D-Cell the ith layer. When the size of
the node degree is increases, the D-Cell is close to 2 exponent index extensions.
D-cell_0 usually contains 3 to 8 unit servers and they are connected through
the micro-switch just as shown in Fig. 2.
|| BCube_1 structure of the cold computing
D-Cell has better fault tolerance capability; the distributed protocol implementation
is close to the shortest path routing. D-Cell also can provide various services
with network capacity which is higher than the traditional tree structure. Additional
D-Cell can increase the extension and in the case of incomplete structure it
also can show the performance. Although DCell_0 is small but DCell can support
the larger number of servers, as known that when DCell_0 contains 6 hosts, DCell_3
can support 3.26 million servers.
The server can realize the D-Cell, link or can be switched under the condition
of serious fault and it achieves good routing performance. D-Cell also has some
shortcomings; first of all, the complete graph connections may lead to enormous
cost, size and the actual link connection and maintenance difficulties. Second,
the uneven distribution of flow at different levels in the DCell, level 0 afford
too much traffic and it will seriously affect the throughput capacity. Finally,
because the DCell adopts the server perform routing, which enlarges the network
latency and the routing protocol is not suitable for find the shortest path
in link failure and the network delay will be larger.
Bcube is DCell modular version, its connection is outside BCube_0 hierarchy
connection and it is realized by using the micro switch as shown in Fig.
3. The number of units in each layer is different from the DCell, if there
has n in the first k DCell_k-1 in DCell layer, The BCube at any level will the
same unit number n, so it can easily obtained that BCube_k has nk+1 servers
and n is the number of servers in a BCube_0. Switches as the medium of connection
make BCube has many redundant paths, this can ensure that the fault tolerant
routing and convenient modular connection and routing speed are faster than
|| The structure of VL2 in the cold computing system
Another Bcube adopts BSR (BCube Source Routing) in selecting path and the
path adaptive protocol, which can develop the network and realize the shortest
parallel paths and reliable data transmission.
Bcube can execute efficiently the One to One, One to All, One to Several, All
to All types of communication without bandwidth limitation and it supports GFS
and MapReduce applications. BCube connection method and the recursive structure
allow data centers to modular construction and the realization is the good.
Under the condition of incomplete structure the BCube has better performance
than DCell. Experiments show that when given 2048 servers, in Fat-Tree, DCell
and BCube structure are not complete, the network throughput and fault-tolerant
performance of BCube are the best one and this is because the incomplete structure
of BCube adopts the complete structure of switch strategies. The deficiency
of BCube is on the extensibility, in the BCube, when in k = 3, n = 8 (k is BCube
levels, n is BCube_0 server number), it only can support 4096 servers, which
is smaller than the DCell.
As shown in Fig. 4, VL2 is an extensible and flexible data
center network structure, it can be able to support large scale data centers,
provide a balanced high bandwidth communication performance for servers, the
performance isolation between services and layer 2 Ethernet semantics. Semantics
refers to the second tier 2, all of the virtual domain into a unified domain,
at this level all hosts in the same domain. VL2 structurally changed is the
connecting way of layer 3 switches, the use of special virtual layer 2 protocol
implementation; And other structural changes in the physical connection is a
whole. In addition, the said address and routing protocol in the VL2 is even
more important, it is directly related to the realization of virtual layer 2.
In the VL2, the VLB (Valiant Load Balancing) routing is adopted to balance various
runoff, VL2 among each switch is set to the same IP, adopt the way of random
choosing an intermediate switch routing.
This section it put forward the method of building data center network structure
of the snowflake and give out some of their special properties. It call be called
the snowflake structure, because according to the construction of the structure
of the Koch curve and its shape is like the Koch snowflake.
Snowflake structure and its construction method: Snowflake structure
contains two parts: the server and the micro switches. The Salt DCell and BCube
structure, they both are recursive method is adopted to define their structure.
In the snow structure, it adopt the method of recursive definition, also at
level (n-1) if a snowflake structure on the basis of adding, add a level 0 snowflake
structure to constitute the n-tier structure of snowflakes, Several methods
will be described in detailed later.
First of all, it take the fixed structure way in the modular structure, which
will be helpful to the structure of modular connection; Second, when the n-tier
structure is not be extend completely, it expand (n+1) level structure is relatively
easy. At this point, if the system discovers that the n-tier structure did not
extend completely, it can be continued to supplement the n-tier structure of
incomplete without changing (n+1) level structure, it is helpful to the expansion
of the structure. Snowflake structure construction method can be described in
detailed as below.
Level 0 snowflake structure (Snow_0) is composed of a micro switch and a number
of servers. According to data center structure, the number of server is at least
3 and not more than 8. As shown in Fig. 5, which takes the
three servers for example (k = 3) and it connects the three servers with the
micro switch port, which is shown in Fig. 5, if the theoretical
position of the server is changed and the three servers are connected directly
with dotted line. The three dotted lines are not the actual connections of the
servers, it can be called virtual connection, there is no actual connection
and it is just for the convenience.
||The structure of Snow_0 when k = 3 in the cold computing system
Based on Snow_0, the three virtual connections can be canceled which is as
shown in Fig. 5, every virtual connections are disconnected,
namely that adding a 0 level snowflake structure and add the new level 0 structure
of micro switches, respectively and the virtual connection nodes are connected
on both ends. Compare with the virtual connection, this connection is called
real connection, or disconnect the virtual connection rebuilt the actual connection.
Real connection is real connection; here is the newly added structure of the
switch with the original structure.
Server connection: So, the Snow_1 snowflake structure can be obtained,
which is as shown in Fig. 6.
Because it is add disconnect on the virtual connection level 0 snowflakes,
the state of the virtual connection has become real connection at this time
namely that the one virtual connection has been changed into two physical connections.
The special note is that the connection is not just a real connection, it is
reflected by the virtual connection to real connection, the connection reflects
the change of states, so in Snow_0 server which is connected to the switch,
three places are not as a real connection, although it is a real connection,
but there is no state change. Through careful observation, it can be found that,
disconnect virtual connection is to add the structure and there is a difference
between Snow_0 level 0, it lacks a virtual connection and it is not really a
Snow_0 structure. In order to distinguish between the two connections, the later
broken open virtual connection and keep adding and which is lack of a virtual
connection structure of level 0, it is the cell to distinguish Snow_0.
||The data structure of Snow_1 when k = 3 in the cloud computing
In the Snow_1 as shown in Fig. 7, it contains six real connections
and six virtual connections. Snowflake structure in each level, always keep
off the next higher level, which contains all of the real and virtual connections,
if Cell is added, a new advanced structure can be formed.
Snowflake structure performance analysis: In this section, the performances
of the snowflake structure from the network exchange are analyzed. i.e., the
capacity and bottleneck link, time delay, green energy-saving four.
Network exchange capacity: If it is assumed that the bandwidth is AMb/s,
the network width between switch and server is about BMb/s and generally the
A is larger than B, when the bandwidth between the two servers is limited to
the minimum bandwidth value of the path. Snowflake structure can maintain a
low number of switches under the premise of good extensibility. In addition,
in snow flower and the structure, the basic server is connected to the switch
and there are almost no existences of server connecting with the server directly.
So, that it can ensure high data transfer rate. Therefore, snowflake structure
is suitable for the data when there needs frequent communication between nodes.
Bottleneck link: Our bottleneck can be divided into two kinds: Bottleneck
node and path bottleneck.
||The improved Snow_1 structure in the cloud computing system
Bottleneck node refers to a node (server or switch) due to the large data
The more the easier to be the bottleneck node level low. So some server in
the hierarchy can improve for switches; the switches can be replaced with high
throughput, such as signs in the switch, in order to ease the pressure on data
Path bottleneck refers that one and only one path between the two subnets,
when the path is disconnected and the two subnets loses connection to form a
bottleneck; the path can be called path bottleneck. Any two servers will prove,
Snow_n contains parallel paths and at least two and not more than 2n in the
theory, Due to this theory, there is no the problem of path bottleneck in snow
Time delay: RTT (Round Trip Time) is the total delay time from the sender
to send data to and receive from reception to the sending end and the confirmation
the receiver send confirmation immediately after receipt of the data. When between
two points RTT value of the larger population of the time delay between the
two; On the other hand is the time delay between two points is smaller. It can
be seen from the definition of RTT, RTT is proportional to the shortest path
between two points.
||Change of average shortest path length with different failure
rate of link
When the distance between two points a long, RTT is larger; On the other hand,
the RTT is smaller.
So, it can conclude that the longest shortest path to jump (2n+1). Therefore,
delay time is O (n).
SIMULATION AND EXPERIMENT RESULT ANALYSIS
In this section, the simulations are adopted to evaluate the snowflake structure
performance. Simulation program in written in Java language, development tools
is Eclipse 3.5.0, run environment is on Lenovo computer with 4 cores 2.7 GHz
Situation 1: When there is no node failure, the total number of nodes
at different levels and average shortest path length.
Situation 2: When there exist the node failures, When k = 3, the detailed
node is set with failure ratio, the test along with the node effect the increase
of the proportion in the Snow_4 and average shortest path length will be changed.
As it can be seen from the Fig. 8, when a node failure is
kept at 0.05-0.05, the average path length fluctuation is not big and the value
is between 2.6 and 2.7. When k = 3, path of failure rate in the Snow_4 and e
DCell are the contrasted through experimental results as shown in the Fig.
9, it is path failure comparison between Snow_4 and DCell_4.
||The Path failure rate between Snow_1 and DCell_4 under different
node failure rate
||Average shortest path of each node to the node Snow_0 when
there no node fails
It can see from the Fig. 9, the failure path of the Snow
efficiency around DCell fluctuate, the difference is not very big. When under
the no-node failure conditions, in the Table 1, it can be
found that when k = 3, the average distance from the Snow_0 server to the rest
of the nodes is the shortest. As the node failure situation is not considered,
therefore number of inaccessible node is the zero. Through the average shortest
path length, it can found in Snow_4 range, average path can be maintained. Situation
3: When there exists a link fails, Here the link failure refers to the link
between the disconnect switch and server, or switch failure, the failure is
not the server itself.
As can be seen from the Fig. 9, with the increase of link
failure rate, average shortest path length changes are not big, the link failure
rate is within 0.2, average shortest path length between 2.6 to 2.9.
When k = 3, some test of the path of failure rate as the link failure rate
increases are made, in the Snow_4. some comparisons with the DCell results in
the experimental are also given out. As shown in Fig. 10,
it is path failure comparison between Snow_4 with DCell_4. It can be found that
the Snow path failure rate is stable which is always maintained between 0.001
||Path failure of Snow_4 and DCell_4 under same link failure
|| Comparison of characteristic between DCell, BCube and Snow
Finally, the comparison between DCell and BCube Snow are presented, the results
are as shown in the Table 2.
, Yu and Vahdat (2002)
Higgs et al. (1997)
et al. (2003)
and Foster and Kesselman (1997)
have made related research, but the research results are not suitable in the could
computing and the new data structure in cloud computing is very necessary.
In order to meet the scalability in the structure of cloud computing data center
and green energy demand, in the study, the new data center network structure
based on the famous Koch curve, the structure of snowflakes is put forward.
In the structure, the scalability of data centers are considered, under the
premise of guarantee switches with the server, it can realize smaller network
expense in a relatively short average path between nodes within the routing
In my opinion the center data structure such as the snowflake structure of
cloud computing data center and network building methods, can improve the performance
of the cloud computing. And the research is necessary and will enhance the speed
and other parameters of the cloud computing. The security of the proposed structure
in the cloud computing is very well and the security issues is not the research
content of the study, but from the related research progress, it can be found
that it is not the problem.