CN103297356A - Method for optimizing switched network cost of data center - Google Patents

Abstract

The invention discloses a method for optimizing switched network cost of a data center. The method for optimizing the switched network cost of the data center comprises performing accumulation on received data packets through an input stage exchanger to form into a flow matrix B; setting up a compromise relationship between the data packet delay and the number of core exchangers through an ADTPT algorithm; computing the least number of the core exchangers under the given delay; and computing minimum cost of a switched network of the data center. The method for optimizing the switched network cost of the data center has the advantages of being flexible in design and increasing expandability of a network switching design of the data center due to the facts that the delay and the core exchangers are designed in a comprise mode and the number of the core exchangers is reduced through suitable delay increasing; improving the design for the network switching cost due to the fact that reconfiguration costs of the core exchangers are incorporated into the total cost of the network switching; and achieving batch scheduling in a data center switched network which satisfies a delay constraint condition and improving a batch scheduling process of multilevel switching due to the fact that single exchanger scheduling of a flow matrix decomposition algorithm is expanded into scheduling of a plurality of exchangers.

Description

A kind of method of optimizing data center's switching network cost

Technical field

The present invention relates to a kind of method of optimizing data center's switching network cost.

Background technology

Along with the fast development of cloud service, data center has been subjected to paying close attention to widely as service supporting platform.For solve the energy consumption that data center's inner exchanging network design faces big, inner connect complicated, cost is high, switching performance is difficult to problems such as assurances, adopts light to exchange the main direction that the novel framework of data center that is connected with light has become technical development.The characteristics of these schemes are to adopt the OpenFlow interchanger as ToR(or Aggregation) interchanger, utilize the crossbar optical exchanger to realize core switch simultaneously, and in data center's switching network, adopt high speed fibre to connect, thereby simplified switching network structure, strengthened the extensibility of system.

Yet, the Clos framework that this multistage exchange forms, in order to realize clog-free exchange, the number needs of intergrade interchanger (being core switch) will increase along with the increase of edge switcher port number in a particular manner.Such as, rearrangeable clog-free in order to realize, the number that need satisfy the intergrade interchanger must not be less than the port number of input stage interchanger.This quantity to the required core switch in data center has been made very strong restriction, thereby has restricted the cost of data center's switching network.

At above-mentioned light packet switching system, adopt packet switch (Packet switching) pattern of lot size scheduling (Batch scheduling) to improve the resource utilization of switching network and the switching performance of system usually.In the lot size scheduling process, packet earlier through after a while accumulation, and then connects unified exchange the under the configuration by interchanger in certain exchange in input-buffer.At the light packet switching system of single interchanger, Bin Wu has provided the dispatching method that decomposes based on traffic matrix.According to its ADAPT algorithm that provides, at each traffic matrix of going and reaching row and all being no more than r * r of nA, the ADAPT algorithm can resolve into N to it _S(r ²-2r+2〉N _SR) individual transposed matrix { P _k(N _S〉=k 〉=1), a kind of configuration of each transposed matrix respective exchanger, and the timeslot number of keeping is:

Yet the ADAPT algorithm only is applied to the scheduling of single interchanger at present; In addition, the major issue that above-mentioned employing crossbar optical exchanger faces as core switch is that the expense of reshuffling of system (mainly comprises core switch and switching the required time overhead of interconnection configuration, and carry out between each interchanger the synchronous time of consuming), continuous expansion along with data center's scale, the lifting of switching network complexity, reshuffle overhead issues and will seem more and more outstanding, yet it is current at the optimized design studies of data center's inner exchanging network cost, all the time the expense of reshuffling of having ignored system, it is not too reasonable to design.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, a kind of method of optimizing data center's switching network cost is provided, it is reshuffled expense with interchanger and includes within the limit of consideration of cost optimization design, the ADAPT algorithm is expanded to multistage switching network, finish traffic matrix decomposition and exchange scheduling, and then realize that the compromise between packet time-delay and the core switch quantity designs, make data center's switching network cost optimization, also broken through simultaneously in the existing design constraint to core switch quantity.

The objective of the invention is to be achieved through the following technical solutions: a kind of method of optimizing data center's switching network cost, it comprises the steps:

S1: the input stage interchanger accumulates the packet of receiving, forms the traffic matrix B of a r * r;

S2: utilize the ADTPT algorithm to set up tradeoff formula between packet time-delay and the core switch quantity, the step of setting up the tradeoff formula is as follows:

S21: adopt the ADTPT algorithm that traffic matrix B is resolved into N _SIndividual transposed matrix P _kWeighted sum;

S22: calculate packet exchange time-delay A with the tradeoff formula between the core switch quantity m;

S3: calculate the minimum core switch number under the given time-delay, computational methods are as follows:

S31: ask for about N at core switch quantity m according to the tradeoff formula _SThe second order local derviation;

S32: find unique N _SMake m obtain minimum value;

S4: the minimum cost of calculated data center switching network, computational methods are as follows:

S41: the conversion cost of definition unit time-delay is τ, and then total cost is intergrade interchanger cost and time-delay cost sum, that is: C=m+ τ A;

S42: according to the functional relation between m and the A, obtain the function of C and A;

S43: ask for C about the second derivative of A;

S44: calculate time-delay border A and required core switch number N _S, and then obtain the switching network minimum cost C of data center.

Further, described packet exchange time-delay A with the tradeoff formula result between the core switch quantity m is:

$A=\frac{{\mathrm{\δN}}_S}{m-\frac{N_S}{N_S-r}}$ And $m=\frac{{\mathrm{\δN}}_S}{A}+\frac{N_S}{N_S-r};$

Wherein, δ is reshuffling the overhead time of core switch.

Further, the minimum value that obtains of described core switch quantity m is:

$m=\frac{{\mathrm{\δN}}_S}{A}+\frac{N_S}{N_S-r}{|}_{N_S=r+\sqrt{\frac{\mathrm{rA}}{\mathrm{\δ}}}}={(1+p)}^2;$

Wherein, $p=\sqrt{\frac{\mathrm{\δr}}{A}}.$

Further, the functional relation of described C and A is:

$C={(1+p)}^2+\mathrm{\τA}={(1+\sqrt{\frac{\mathrm{\δr}}{A}})}^2+\mathrm{\τA}.$

The invention has the beneficial effects as follows:

(1) the packet time-delay designs with the compromise of core switch quantity, can delay time to reduce the quantity of required core switch by suitable increase, flexible design, and strengthened the extensibility that data center's switching network designs;

(2) with the total cost that expense is brought data center's switching network into of reshuffling of core switch, it is more reasonable to design, and perfect switching network cost designs a model;

(3) the traffic matrix decomposition algorithm is expanded to the scheduling of a plurality of interchangers from the scheduling of single interchanger, realized the lot size scheduling in the data center's switching network under the satisfied time-delay constraints, the perfect lot size scheduling process of multistage exchange.

Description of drawings

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is the structure chart of three grades of switching networks.

Embodiment

Below in conjunction with accompanying drawing technical scheme of the present invention is described in further detail, but protection scope of the present invention is not limited to the following stated.

As shown in Figure 2, three grades of Clos (n, m, r) switching network framework, the input stage interchanger that r n * m arranged in this network, the output interchanger of m r * r intergrade interchanger (being core switch) and r m * n all is communicated with between any two ends mouth of adjacent two-stage interchanger (input port of the output port of previous stage, back one collection).

As shown in Figure 1, a kind of method of optimizing data center's switching network cost, it comprises the steps:

S1: the input stage interchanger accumulates the packet of receiving, the accumulation cycle is each time slot of A, forms the traffic matrix B of a r * r.Any row of B or arbitrarily row element and all be no more than nA.Owing to adopted the light connection, the M that the inner packet transmission rate of core switch is the individual server packet transmission rate is (M 〉=n), suppose M=n here doubly.

S2: utilize the ADTPT algorithm to set up tradeoff formula between packet time-delay and the core switch quantity, the step of setting up the tradeoff formula is as follows:

S21: adopt the ADTPT algorithm that traffic matrix B is resolved into N _SIndividual transposed matrix P _kWeighted sum;

The overhead time of reshuffling of supposing intergrade interchanger (core switch just) is δ, in order to make the buffer of ToR interchanger store data bag not collapse under pressure, the core stage interchanger must be finished the exchange of all packets in A time slot, so can obtain following restriction relation:

Wherein, m is the number of core switch,

For interchanger is configured to P _kThe time timeslot number that need keep, In above-mentioned restriction relation formula

Be the scheduling total length in A the time slot, this total length is on average born by m core switch.

S22: calculate packet exchange time-delay A with the tradeoff formula between the core switch quantity m;

Further above-mentioned restriction relation formula is carried out abbreviation, can obtain packet exchange time-delay A with the tradeoff formula result between the core switch quantity m:

$A=\frac{{\mathrm{\δN}}_S}{m-\frac{N_S}{N_S-r}}$ With $m=\frac{{\mathrm{\δN}}_S}{A}+\frac{N_S}{N_S-r};$

Wherein, δ is reshuffling the overhead time of core switch.When the A value is more big, packet exchange time-delay is also just more big, and required core switch is also just more few.

S3: calculate the minimum core switch number under the given time-delay, computational methods are as follows:

S31: ask for about N at core switch quantity m according to the tradeoff formula _SThe second order local derviation because given δ, r and A then obtain:

$\frac{d^{2}m}{d{N_S}^2}=\frac{2r}{{(N_S-r)}^3}>0;$

In order to realize that switching performance guarantees that N must be arranged _SR.Therefore, m is N _SConcave function, and can find unique N _SMake m obtain minimum value.Formula below the order is set up:

$\frac{\mathrm{dm}}{{\mathrm{dN}}_S}=\frac{\mathrm{\δ}}{A}-\frac{r}{{(N_S-r)}^2}=0;$

Can obtain:

$N_S=r+\sqrt{\frac{\mathrm{rA}}{\mathrm{\δ}}};$

N wherein _SBe necessary for integer, therefore in practice should value

S32: find unique N _SMake m obtain minimum value;

$m=\frac{{\mathrm{\δN}}_S}{A}+\frac{N_S}{N_S-r}{|}_{N_S=r+\sqrt{\frac{\mathrm{rA}}{\mathrm{\δ}}}}={(1+p)}^2;$

Wherein, $p=\sqrt{\frac{\mathrm{\δr}}{A}}.$

S4: the minimum cost of calculated data center switching network, computational methods are as follows:

S41: the conversion cost of definition unit time-delay is τ, and then total cost is intergrade interchanger cost and time-delay cost sum, that is: C=m+ τ A;

S42: according to the functional relation between m and the A, obtain the function of C and A, namely

$C={(1+p)}^2+\mathrm{\τA}={(1+\sqrt{\frac{\mathrm{\δr}}{A}})}^2+\mathrm{\τA};$

S43: ask for C about the second derivative of A:

$\frac{d^{2}C}{d{A}^2}=\frac{1}{A^3}(\frac{3\sqrt{\mathrm{\δrA}}}{2}+2\mathrm{\δr})>0;$

Therefore, C is the concave function of A, and C has unique minimum value:

$\frac{\mathrm{dC}}{\mathrm{dA}}=-\frac{\mathrm{\δr}}{A^2}-\sqrt{\frac{\mathrm{\δr}}{A^3}}+\mathrm{\τ}=-\frac{1}{\mathrm{\δr}}(p^4+p^3-\mathrm{\τ\δr})=0.$

S44: according to above-mentioned function, can find the p value by the method for numerical solution, calculate time-delay border A and required core switch number N _S, and then obtain the switching network minimum cost C of data center.

Claims (4)

1. method of optimizing data center's switching network cost, it is characterized in that: it comprises the steps:

S1: the input stage interchanger accumulates the packet of receiving, forms the traffic matrix B of a r * r;

S2: utilize the ADTPT algorithm to set up tradeoff formula between packet time-delay and the core switch quantity, the step of setting up the tradeoff formula is as follows:

S21: adopt the ADTPT algorithm that traffic matrix B is resolved into N _SIndividual transposed matrix P _kWeighted sum;

S22: calculate packet exchange time-delay A with the tradeoff formula between the core switch quantity m;

S3: calculate the minimum core switch number under the given time-delay, computational methods are as follows:

S31: ask for about N at core switch quantity m according to the tradeoff formula _SThe second order local derviation;

S32: find unique N _SMake m obtain minimum value;

S4: the minimum cost of calculated data center switching network, computational methods are as follows:

S41: the conversion cost of definition unit time-delay is τ, and then total cost is intergrade interchanger cost and time-delay cost sum, that is: C=m+ τ A;

S42: according to the functional relation between m and the A, obtain the function of C and A;

S43: ask for C about the second derivative of A;

S44: calculate time-delay border A and required core switch number N _S, and then obtain the switching network minimum cost C of data center.

2. a kind of method of optimizing data center's switching network cost according to claim 1 is characterized in that: described packet exchange time-delay A with the tradeoff formula result between the core switch quantity m is:

$A=\frac{{\mathrm{\δN}}_S}{m-\frac{N_S}{N_S-r}}$ And $m=\frac{{\mathrm{\δN}}_S}{A}+\frac{N_S}{N_S-r};$

Wherein, δ is reshuffling the overhead time of core switch.

3. a kind of method of optimizing data center's switching network cost according to claim 1, it is characterized in that: the minimum value that described core switch quantity m obtains is:

$m=\frac{{\mathrm{\δN}}_S}{A}+\frac{N_S}{N_S-r}{|}_{N_S=r+\sqrt{\frac{\mathrm{rA}}{\mathrm{\δ}}}}={(1+p)}^2;$

Wherein, $p=\sqrt{\frac{\mathrm{\δr}}{A}}.$

4. a kind of method of optimizing data center's switching network cost according to claim 1, it is characterized in that: the functional relation of described C and A is:

$C={(1+p)}^2+\mathrm{\τA}={(1+\sqrt{\frac{\mathrm{\δr}}{A}})}^2+\mathrm{\τA}.$

Images