CN105207944A - FAST TCP based transmission control method - Google Patents

Abstract

The invention discloses a FAST TCP based transmission control method. A data control module, a window control module and an estimation module are included, and the estimation module is used to calculate a congestion avoidance threshold. The method comprises the following steps that S1) the congestion avoidance threshold and the amount of data packages transmitted currently of a client are obtained; and S2) the amount of the data packages transmitted currently of the client is compared with the congestion avoidance threshold, if the amount is not lower than the congestion avoidance threshold, the estimation module sends congestion feedback information to the window control module, and the window control module reduces the amount of transmission data packages according to the congestion feedback information to avoid package loss. The method has the advantages that the network transmission scale is monitored, so that parameters can be adjusted according to different scales, adaption can be realized in the different scales, and high throughput and stable transmission can be ensured always.

Description

Based on the transfer control method of FAST TCP

Technical field

This technology belongs to cloud computing and network transmission technology field, is specifically related to a kind of transfer control method based on FASTTCP.

Background technology

The topological structure of data center, through exploration for many years, has had very large change.Fig. 1 shows the topological structure of conventional data center network.In this topological structure, scope (ToR) the switch Access Layer of frame top formula provides connection for rack-mounted server.(AggregationSwitch) traffic forwarding that will send from multiple Access Layer (TOR) is exchanged to core layer in the gathering of each Guinier-Preston zone (being also sometimes referred to as distribution layer).Each ToR switch is connected to the redundancy of multiple aggregation switch.Switch is passed through core route CoreRouters by core layer) be connected to the Internet.A special case is two smooth etale topologies, only uses two layer switch.

Clos topology is a kind of topological structure based on multistage switches, and each switch level is connected with all switches of next switch level, considerably increases the diversity in path like this.Fig. 2 illustrates a kind of typical three layers of Clos topology

Fat tree is a kind of special Clos topological structure.It is a kind tree structure, and as shown in Figure 3, this topological structure is made up of the switch of k port, and they have k tie point, and there is two-layer (polymer layer and marginal layer) k/2 switch at each tie point place, and every (k/2) ²individual core switch has 1 port and this k tie point link.Core layer switch is connected with polymer layer switch, and edge once switch be connected with Tor switch.This strategy has been traded off complexity and route diversity.

Rising in recent years virtual data center network, virtual data center network is the virtual link formed between some virtual resources (virtual machine, virtual switch, virtual router etc.), as shown, a typical data center can suppose multiple virtual data center.Although introduce virtual concept, the utilization of resources can have Fig. 4 to be erected at virtual data center on typical data center and data center network effect more flexibly, user also can shield the matters of physical layer, but data center network problem still objective reality, must solve.

One of problem that Transmission Control Protocol brings to data center is exactly TCPIncast phenomenon, and first we will define data center network environment here.The following condition of data center network environment general satisfaction:

A) server has high-performance CPU and larger internal memory;

B) high-performance network interface card is possessed, and suitable operating system;

C) gigabit Ethernet (or 10Gb/sec), by a switch or other networkings, topological structure composition data center, more multi-exchange may bring benefit, but definition has no particular limits.

D) exist in a large number one to one, many-one and one-to-many communication pattern.This type of transmission mode is present in other many data-intensive Distributed Application such as cluster-based storage and Hadoop.

Under simple bandwidth environment, such as use scp by the data Replica on five machines to a destination host, each source machine can export more than 900Mbps.If each sender sends data with the speed of 900Mbps to receiving terminal, 1G Ethernet switch soon can only discarded part divided data bag.

But this situation is readily appreciated that: all machines can not send the data of more than 900Mbps simultaneously, because receiving terminal has the bandwidth upper limit, switch will abandon some packets, and then each TCP connects and will reduce packet sending speed.Finally, each sender will reach the stable state approximately sending 200Mbps thus fill up bottleneck link.

In fact we can run into more difficult bandwidth problem, if there be larger trooping, there will be a large amount of server simultaneously to the situation of a server transmission data, then may have the flow " microexplosion " being sent to single cluster member.Switch can not cushion all " microexplosion frame ", therefore have partial frame to be lost, the throughput collapse in the gap in short-term that data display below " microexplosion " causes.Then, the congestion control mechanism of TCP starts to intervene, process packet loss and time out event, slows down window growth and even reduces window.

" Incast " is a kind of communication pattern, and in the data under heart network environment, it causes primarily of the feature of the congestion avoidance algorithm of Transmission Control Protocol.Figure 5 shows that typical TCPIncast transmission mode scene.In this mode, client is connected to the multiple servers of data center by a switch.Client sends request, and data are transferred to client in the mode of many-to-one by bottleneck link by switch from one or more server.Client is with larger logical block (as 1MB) request msg, and the data block of reality is then stored in multiserver with less data block (as 32KB) form, is called the request unit (SRU) of server.Each takes turns in transmission, and client sends request of data to all servers storing requested date, when the data of Servers-all all successfully receive (i.e. requested date block Successful transmissions), just carries out data block request next time.

Define following three conditions of Incast phenomenon demand fulfillment rigorously:

1. network environment: high bandwidth (kilomega network) low delay and less switch buffer memory

2. synchronization blocks transmission mode, namely often takes turns transmission data block, and all data blocks receive the rear next round that just enters and transmit

3. less data block; (maximum 256K, second can end of transmission in 2 millis under kilomega network environment for a blocks of data)

Along with the increase of concurrent sender's quantity, volume of transmitted data may exceed switch cache size, causes data-bag lost, and the Retransmission timeout of the thing followed.Thus throughput collapse (sharply reducing) may be caused, as shown in Figure 6, we are referred to as TCPIncast phenomenon.

TCPIncast problem may appear in the middle of many typical data-center applications.Such as, in cluster storage system, when the request of memory node to data responds; In network search procedure, many working machines almost respond search inquiry simultaneously; Or in the process of MapReduce batch process operation, when " shuffle " stage, intermediate data was transferred to Reducer by multiple Mapper.Because existence range is extensive, make a very bad impression to network efficiency, TCPIncast problem has caused the attention of many researchers.Researcher proposes various possible solution at stage construction, mainly comprises link layer, transport layer and application layer solution.In these schemes, part has higher efficiency, but cost is also high, and segmentation scheme cost is low, but produces little effect.

Summary of the invention

The present invention is intended at least one of solve the problems of the technologies described above.

For this reason, one object of the present invention is to propose a kind of transfer control method based on FASTTCP.

To achieve these goals, the embodiment of a first aspect of the present invention discloses a kind of transfer control method based on FASTTCP, comprising: data control block, and described data control block is used for controls transfer packet; Window controlling module, described window controlling module is for controlling the transmission quantity of the packet of transmitting terminal; And estimation module, the average delay that described estimation module is used for packet in the time delay queue of bandwidth according to the Packet Generation frequency of transmitting terminal, transmission link, the buffer memory capacity of switch, described switch calculates Congestion Avoidance threshold value; Said method comprising the steps of: S1: the transmission quantity obtaining described Congestion Avoidance threshold value and described client current data packet; S2: the transmission quantity of current data packet described in described client and described Congestion Avoidance threshold value are compared, if the transmission quantity of described client current data packet is not less than described Congestion Avoidance threshold value, described estimation module sends congestion feedback information to described window controlling module, and described window controlling module reduces the transmission quantity of packet to avoid packet loss according to described congestion feedback information.

According to the transfer control method based on FASTTCP of the embodiment of the present invention, scale can be transmitted by monitor network, carry out parameter adjustment for different scales, reach the self adaptation under different scales, ensure high-throughput and stable transmission performance all the time.

In addition, the transfer control method based on FASTTCP according to the above embodiment of the present invention, can also have following additional technical characteristic:

Further, the computing formula of described Congestion Avoidance threshold value is as follows:

$w\←min\{2w,(1-\mathrm{\γ})w+\mathrm{\γ}(\frac{baseRTT}{RTT}w+\mathrm{\α}(w,qdelay\left)\right)\}$

Wherein, w is the transmission quantity of the described current data packet of described transmitting terminal, γ ∈ (0,1], baseRTT is minimum RTT, the qdelay of packet in described switch time delay queue is point-to-point average queue delay, and α is correction factor,

$\mathrm{\α}=\left\{\begin{array}{cc}{\mathrm{\α}}_0& q\≤q^{*}\\ {\mathrm{\α}}_1& q>q^{*}\end{array}\right.,$

Wherein, q is queue time delay, α ₀, α ₁, q ^*be constant.

Additional aspect of the present invention and advantage will part provide in the following description, and part will become obvious from the following description, or be recognized by practice of the present invention.

Accompanying drawing explanation

Above-mentioned and/or additional aspect of the present invention and advantage will become obvious and easy understand from accompanying drawing below combining to the description of embodiment, wherein:

Fig. 1 is switch-layer topology structural representation in prior art;

Fig. 2 is Clos topological structure schematic diagram of the prior art;

Fig. 3 is fat tree topology structural representation of the prior art;

Fig. 4 is virtual data center of the prior art and data center network schematic diagram;

Fig. 5 is TCPIncast transmission mode schematic diagram of the prior art;

Fig. 6 is the schematic diagram of TCPIncast of the prior art;

Fig. 7 is the flow chart of one embodiment of the invention.

Embodiment

Be described below in detail embodiments of the invention, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar element or has element that is identical or similar functions from start to finish.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.

In describing the invention, it will be appreciated that, term " " center ", " longitudinal direction ", " transverse direction ", " on ", D score, " front ", " afterwards ", " left side ", " right side ", " vertically ", " level ", " top ", " end ", " interior ", orientation or the position relationship of the instruction such as " outward " are based on orientation shown in the drawings or position relationship, only the present invention for convenience of description and simplified characterization, instead of indicate or imply that the device of indication or element must have specific orientation, with specific azimuth configuration and operation, therefore limitation of the present invention can not be interpreted as.In addition, term " first ", " second " only for describing object, and can not be interpreted as instruction or hint relative importance.

In describing the invention, it should be noted that, unless otherwise clearly defined and limited, term " installation ", " being connected ", " connection " should be interpreted broadly, and such as, can be fixedly connected with, also can be removably connect, or connect integratedly; Can be mechanical connection, also can be electrical connection; Can be directly be connected, also indirectly can be connected by intermediary, can be the connection of two element internals.For the ordinary skill in the art, concrete condition above-mentioned term concrete meaning in the present invention can be understood.

With reference to description below and accompanying drawing, these and other aspects of embodiments of the invention will be known.Describe at these and in accompanying drawing, specifically disclose some particular implementation in embodiments of the invention, representing some modes of the principle implementing embodiments of the invention, but should be appreciated that the scope of embodiments of the invention is not limited.On the contrary, embodiments of the invention comprise fall into attached claims spirit and intension within the scope of all changes, amendment and equivalent.

Below in conjunction with accompanying drawing, the transfer control method based on FASTTCP according to the embodiment of the present invention is described.

C-FAST agreement adjusts congestion window, to maintain Stability parameter by detecting queue delay; We think that 0 queue length is dangerous, prove that bottleneck link is occupied full, and when 0 queue length, we cannot judge whether bottleneck link is fully used because there is queue to exist; Meanwhile, in the data in the heart, the yardstick of RTT is very little, and queue delay is then less, and we select directly based on queue delay, and the congestion avoidance policy maintaining the FAST agreement of >0 Stability parameter is as the basis of dealing with problems.

This agreement is based on FAST Protocol Design.FAST agreement uses in backbone network, and the primary difference of backbone network and data center is as follows:

RTT (round-trip delay), backbone network is at about 200ms, and data center is at 10us ~ 100us, thus the order of magnitude of qdelay also drops to us level from ms level;

Buffer size, in backbone network, the buffer size of router is about 100Mb, and data center switchboard buffering area is at below 1Mb.

Scheme due to us adopts the congestion avoidance algorithm of FAST agreement, but FAST agreement is necessarily inapplicable in the heart in the data, therefore we need to algorithm adjustment, the program can successfully to be applied based on above 2 differences in the data in the heart, it should be noted that, we carry out for kilomega network in current design.

Congestion Avoidance threshold value refers to that the quick propagation process of window terminates to enter the threshold value of Congestion Avoidance, and different agreements has different standards.Transmission Control Protocol take window size as standard, enters congestion avoidance procedure when congestion window size exceedes this threshold value by slow turn-on.

In FAST agreement, take queue delay as standard, when queue delay is greater than this threshold value, enter Congestion Avoidance.The physical significance of Congestion Avoidance threshold value thinks that link is close to saturated point.Because router cache is general comparatively large, thus waiting list is longer, therefore network delay is comparatively large in backbone network, and therefore this threshold value of FAST agreement is also larger.In the data under heart network environment, switch buffer memory is often less and queue delay is also corresponding less, therefore this threshold value needs to adjust.Because we will make there is queue in buffering area, fully loaded to ensure bottleneck link, will ensure not occur buffer overflow before slow turn-on terminates, therefore Congestion Avoidance threshold value Mi_threshold need meet following condition simultaneously:

SU/C<Mi_threshold<B/C

Wherein SU is bag size, and B is buffer size, and C is link bandwidth.For current application kilomega network more widely, if buffersize (buffer size) for 64KB, SU be 1KB, then

1us<Mi_threshold<64us

Suggestion allows this threshold value away from buffersize, otherwise is easy to because packet loss appears in accidental cause.In fact a less mi_threshold has been enough to ensure that bottleneck link is fully utilized, and we get here:

Mi_threshold＝20us

α is the correction factor that in FAST agreement, congestion window controls, and is the function of w and q-delay.But in current FAST protocol application, be generally taken as constant.Its physical significance is the queue length that current stream is expected to maintain in the buffer, is therefore also the lower bound of stream congestion window.When α is less, it is comparatively slow that namely link convergence enters equilibrium state, and transmitting terminal window size also receives restriction.When α is larger, queueing delay can be increased again.

In backbone network, because router buffering area is comparatively large, α value is also comparatively large, be defaulted as 200, and, the window size of 200 is necessarily unpractical in the data in the heart.Therefore we must adopt new α.

For providing reference to conceptual design, first we devise emulation experiment to estimate the possible span of α, we use NS2 to carry out emulation platform building, emulation platform achieves the transmission network topology structure of Incast and achieves preliminary solution with test parameter based on FASTTCP.Concrete emulation platform can be introduced below.We conducted the test of parameter alpha, we have two conclusions:

When transmitting terminal number is less, less α can limit throughput.Especially point-to-point do transfer of data time, such restriction is more obvious.The prerequisite that we adopt new scheme to solve Incast problem does not limit normal data transfer, and α=0, the situation of 1 has obviously run counter to this point;

When larger α can make transmitting terminal number large, the minimum value of total queue length is greater than switch buffer memory, thus causes incast phenomenon.Because α is the queue length that each stream maintains at switch place, if therefore α is comparatively large, likely occur that when number of servers increases total queue length exceeds the situation of switch buffer memory.

Therefore the α of constant value cannot meet the demand of point-to-point transmission demand and Incast transmission mode, and in view of above two features, our consideration determines the value of α with step function:

$\mathrm{\&alpha;}=\left\{\begin{array}{cc}{\mathrm{\&alpha;}}_0& N\&le;N^{*}\\ {\mathrm{\&alpha;}}_1& N>N^{*}\end{array}\right.$

Wherein α ₀and α ₁be respectively the value of α during different transmitting terminal quantity, N ^*for the threshold value of transmitting terminal number.

α is the correction factor that in FAST agreement, congestion window controls, and is the function of w and q-delay.But in current FAST protocol application, be generally taken as constant.Its physical significance is the queue length that current stream is expected to maintain in the buffer, is therefore also the lower bound of stream congestion window.When α is less, it is comparatively slow that namely link convergence enters equilibrium state, and transmitting terminal window size also receives restriction.When α is larger, queueing delay can be increased again.As mentioned in summary of the invention, its optimum configurations obeys following formula.

$\mathrm{\&alpha;}=\left\{\begin{array}{cc}{\mathrm{\&alpha;}}_0& N\&le;N^{*}\\ {\mathrm{\&alpha;}}_1& N>N^{*}\end{array}\right.$

N ^*the threshold value switching α value, but N ^*concrete value be not a definite numerical value, N in fact ^*have a safe range, within the scope of this, value all can ensure that transmitting terminal number hour link uses saturated and do not occur Incast phenomenon when transmitting terminal number is large.

N ^* _minneed to make the capacity of bottleneck link under minimum window condition to be occupied full.Because work as N>N ^* _mintime α get smaller value, if N ^*value is less than N ^* _minthen may there is the situation that link performance is not occupied full.

Transmitting terminal packet sending speedfor Nw _min, link transmission ratefor CD=12.5pkt, wherein C=1Gbps (kilomega network), D are that network round-trip postpones, D=100us.Therefore N ^* _minmeet following formula:

$N_{\min }^{*}=\min \left\{{\mathrm{Nw}}_{\min }>CD=12.5pkt\right\}=7$

N ^* _maxwhen α need be made to get higher value, switch buffer memory does not overflow, therefore need meet:

$N_{max}^{*}{\mathrm{\&alpha;}}_0=B$

Wherein B is switch buffer memory (bag quantity), and the minimum value of B is 32 packets under normal circumstances.Therefore N ^*safe range be:

N ^*∈[7,B/α ₀],B>32

We still need to continue to determine α ₀n could be determined ^*safe range.In fact, can be learnt by result below, N ^*have a wider safe range, in the process of practical application, we can test the possible value in safe range and choose the numerical value of better performances.We determine respectively below

α ₀should be enough large, make flow transmission reach balanced, and bottleneck link should be made to reach capacity.But we get relatively little α at suggestion here ₀.Because mention above, α ₀relate to N ^*safe range size, less α ₀larger safe range can be ensured, as α ₀the excessive safe range that even makes disappears.Here we get

α ₀＝2

When transmitting terminal quantity is larger, we need less α to ensure not occur Incast phenomenon in transmitting terminal increase process.Therefore α ₁need meet:

Nα ₁<B

α ₁=0 looks it is good selection, but this may limit the transmission rate of single transmitting terminal, the numerical value of a less stable, if all transmitting terminals are all devoted to the queue length at switch place to maintain 0, we cannot judge whether link is fully used, and therefore this numerical value is dangerous.Therefore we select α ₁=1.

We obtain now the relation sending terminal number and α, but the transport layer protocol of transmitting terminal unablely knows currently have how many transmitting terminals to send data to receiving terminal, therefore we need the independent variable of this step function to be converted into the parameter that host-host protocol can be found out.The quantity of transmitting terminal can be estimated by queue delay, due to N ^*there is looser safe range, therefore estimate that the criterion that transmitting terminal quantity draws is safe with queue delay.The relation of N and queue delay qdelay is as follows:

N＝qdelay·C/α

Therefore the value formula of α is converted into:

$\mathrm{\&alpha;}=\left\{\begin{array}{cc}{\mathrm{\&alpha;}}_0& q\&le;q^{*}\\ {\mathrm{\&alpha;}}_1& q>q^{*}\end{array}\right.,q^{*}=N\mathrm{\&alpha;}/C,{\mathrm{\&alpha;}}_0=2,{\mathrm{\&alpha;}}_1=1$

Wherein, q is queue time delay, and we can draw q ^*the safe range of value:

q ^*＝25μs

Namely α value formula is in actual applications:

$\mathrm{\&alpha;}=\left\{\begin{array}{cc}2& q\&le;25\mathrm{\&mu;}s\\ 1& q>25\mathrm{\&mu;}s\end{array}\right.$

Wherein queue delay q is that each RTT can calculate, on the algorithm basis of primary FASTTCP, we do not require that system finds out new parameter, but utilize existing data to carry out being applicable to the redesign of data center (comprising Incast transmission mode) to key parameter α, the benefit done like this is more convenient when real system is implemented, and is more conducive to Project Realization in the future and extensive use.

In addition, other formation of the transfer control method based on FASTTCP of the embodiment of the present invention and effect are all known for a person skilled in the art, in order to reduce redundancy, do not repeat.

In the description of this specification, specific features, structure, material or feature that the description of reference term " embodiment ", " some embodiments ", " example ", " concrete example " or " some examples " etc. means to describe in conjunction with this embodiment or example are contained at least one embodiment of the present invention or example.In this manual, identical embodiment or example are not necessarily referred to the schematic representation of above-mentioned term.And the specific features of description, structure, material or feature can combine in an appropriate manner in any one or more embodiment or example.

Although illustrate and describe embodiments of the invention, those having ordinary skill in the art will appreciate that: can carry out multiple change, amendment, replacement and modification to these embodiments when not departing from principle of the present invention and aim, scope of the present invention is by claim and equivalency thereof.

Claims (2)

1. based on a transfer control method of FASTTCP, it is characterized in that, comprising:

Data control block, described data control block is used for controls transfer packet;

Window controlling module, described window controlling module is for controlling the transmission quantity of the packet of transmitting terminal; And

Estimation module, the average delay that described estimation module is used for packet in the time delay queue of bandwidth according to the Packet Generation frequency of transmitting terminal, transmission link, the buffer memory capacity of switch, described switch calculates Congestion Avoidance threshold value;

Said method comprising the steps of:

S1: the transmission quantity obtaining described Congestion Avoidance threshold value and described client current data packet;

S2: the transmission quantity of current data packet described in described client and described Congestion Avoidance threshold value are compared,

If the transmission quantity of described client current data packet is not less than described Congestion Avoidance threshold value, described estimation module sends congestion feedback information to described window controlling module, and described window controlling module reduces the transmission quantity of packet to avoid packet loss according to described congestion feedback information.

2. the transfer control method based on FASTTCP according to claim 1, is characterized in that, the computing formula of described Congestion Avoidance threshold value is as follows:

$w\&LeftArrow;min\{2w,(1-\mathrm{\&gamma;})w+\mathrm{\&gamma;}(\frac{baseRTT}{RTT}w+\mathrm{\&alpha;}(w,qdelay\left)\right)\}$

Wherein, w is the transmission quantity of the described current data packet of described transmitting terminal, γ ∈ (0,1], baseRTT is minimum RTT, the qdelay of packet in described switch time delay queue is point-to-point average queue delay, and α is correction factor,

$\mathrm{\&alpha;}=\left\{\begin{array}{cc}{\mathrm{\&alpha;}}_0& q\&le;q^{*}\\ {\mathrm{\&alpha;}}_1& q>q^{*}\end{array}\right.,$

Wherein, q is queue time delay, α ₀, α ₁, q ^*be constant.