CN114938350B - Congestion feedback-based data stream transmission control method in lossless network of data center - Google Patents
Congestion feedback-based data stream transmission control method in lossless network of data center Download PDFInfo
- Publication number
- CN114938350B CN114938350B CN202210676107.2A CN202210676107A CN114938350B CN 114938350 B CN114938350 B CN 114938350B CN 202210676107 A CN202210676107 A CN 202210676107A CN 114938350 B CN114938350 B CN 114938350B
- Authority
- CN
- China
- Prior art keywords
- congestion
- output port
- ecn
- data
- queue
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L47/00—Traffic control in data switching networks
- H04L47/10—Flow control; Congestion control
- H04L47/12—Avoiding congestion; Recovering from congestion
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L47/00—Traffic control in data switching networks
- H04L47/10—Flow control; Congestion control
- H04L47/26—Flow control; Congestion control using explicit feedback to the source, e.g. choke packets
- H04L47/263—Rate modification at the source after receiving feedback
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/50—Reducing energy consumption in communication networks in wire-line communication networks, e.g. low power modes or reduced link rate
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses a data stream transmission control method based on congestion feedback in a lossless network of a data center, which comprises the following steps: detecting the queue length of a data packet of an output port of the switch, and judging the congestion state of the output port; if in the congestion-free state, directly forwarding the data packet; if the congestion state is in a continuous congestion state, starting an ECN marking function, and sending the data packet through a target output port after ECN marking so as to control the sending rate according to the ECN marking; if the congestion state is in the burst congestion state, the ECN marking function is disabled until the length of the queue drops by a specified threshold value, the data packet is forwarded, and meanwhile, a congestion notification message CNM is generated and a sender of the congestion flow is directly notified, so that the sending rate is controlled according to the congestion notification message CNM. The method can quickly feed back the sudden congestion, avoids the PFC triggering and the queue head blocking caused by PFC, and has the advantages of simple implementation method, low control implementation cost, high efficiency and the like.
Description
Technical Field
The invention relates to the technical field of lossless networks of data centers, in particular to a data stream transmission control method based on congestion feedback in a lossless network of a data center.
Background
RoCE (RDMA over converged ethernet) technology based on remote direct memory access (remote direct memory access, RDMA) is widely deployed in ethernet datacenter networks (data center network, DCN) in order to reduce data center internal network transmission latency and improve network throughput. However, in the RDMA transmission process, even a single packet is lost, the network throughput is greatly reduced, so that the flow completion time is greatly increased, and the performance of the application service is seriously damaged.
To ensure efficient and reliable RDMA data transfer, data center Ethernet networks are widely deployed with priority-based flow control (PFC) mechanisms to prevent cache overflows. The PFC mechanism is a hop-by-hop flow control mechanism based on ports, and when the length of a queue of an inlet port of a switch exceeds a pause threshold of PFC, a PFC pause message is sent to an upstream switch, and data transmission of an outlet port related to the upstream switch is paused; and when the length of the queue of the ingress port is reduced to be smaller than the recovery threshold value of PFC, sending a PFC recovery message to the upstream port, and recovering the data transmission of the PFC recovery message.
However, the port-based PFC suspension/resumption mechanism described above is extremely prone to problems such as queue blocking, congestion spreading, and deadlock. I.e. when a switch has a certain egress port suspended by PFC, this will cause packets in the queue to be blocked to other non-congested egress ports. More seriously, when a certain switch in the network continuously generates congestion, the PFC hop-by-hop flow control mechanism finally enables an upstream switch irrelevant to the congestion to receive PFC pause signals and pause forwarding of data packets, so that queuing delay of the data packets is increased, network throughput rate is reduced, and transmission completion time of a stream is obviously increased.
In order to solve the above problem, some end-to-end transmission protocols at the flow level, such as DCQCN (Data Center Quantized Congestion Notification), TIMELY, HPCC, swift and PCN, are proposed successively, which can effectively relieve network congestion and reduce the triggering times of PFC. However, when the above transmission protocol solves network congestion, the following problems still exist:
1. at least 1 RTT is required to perceive network congestion, and thus multiple RTTs are typically required to converge the rate of the flow to the target rate.
2. Although the continuous congestion caused by long flows can be effectively controlled, the burst congestion caused by burst short flows with extremely short life time is difficult to control.
3. In the case of bursty congestion, even if the end-to-end transmission protocol is deployed, PFC is inevitably triggered, and the negative problem of PFC occurs.
Therefore, how to quickly feed back congestion and avoid PFC triggering, thereby avoiding negative effects of PFC, accelerating completion of flow, and improving application performance and user experience is a problem to be solved.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: aiming at the technical problems in the prior art, the invention provides the congestion feedback-based data stream transmission control method in the lossless network of the data center, which has the advantages of simple implementation method, low cost and high efficiency, and can quickly feed back congestion, avoid PFC triggering and congestion caused by PFC, accelerate the completion of stream, and promote the application performance and user experience.
In order to solve the technical problems, the technical scheme provided by the invention is as follows:
a data stream transmission control method based on congestion feedback in a lossless network of a data center comprises the following steps:
detecting the queue length of a data packet of an output port of the switch, and judging the congestion state of the output port according to the detected queue length;
if the congestion state of the output port is in a congestion-free state, directly forwarding the data packet to the destination output port to be sent to the host end;
if the congestion state of the output port is in a continuous congestion state, starting an ECN marking function, and sending data packets to a host end through a target output port after ECN marking is carried out on specified data packets in a data queue so as to control the sending rate of the host end according to the ECN marking;
if the congestion state of the output port is in a burst congestion state and the data packets of the congestion flow share the input port with the non-congestion flow, the ECN marking function is disabled until the length of the queue drops by a specified threshold value, the data packets are forwarded to the destination output port, and meanwhile, a congestion notification message CNM is generated and the sending end of the congestion flow is directly notified, so that the sending rate of the host end is controlled according to the congestion notification message CNM.
Further, the determining the congestion state of the output port according to the detected queue length includes: when the length of the queue is smaller than a preset ECN marking threshold QECN, judging that the output port is in a congestion-free state; when the length of the queue is between a preset ECN marking threshold QECN and a preset burst congestion feedback threshold QCNM, judging that the output port is in a continuous congestion state; and when the length of the queue exceeds a preset burst congestion feedback threshold QCNM, judging that the output port is in a burst congestion state.
Further, if the congestion state of the output port is in the continuous congestion state, ECN marking is performed on the data packets exceeding the preset ECN marking threshold QECN in the data queue.
Further, if the congestion state of the output port is in the burst congestion state, judging whether the packet shares the input port with the non-congestion flow, if so, disabling the ECN marking function until the length of the queue falls below the preset ECN marking threshold QECN, and then restarting the ECN marking function, otherwise, directly forwarding the data packet.
Further, the value of the burst congestion feedback threshold QCNM is in the range [ Q ] ECN ，max(Q ECN ,Q PFC /M-3dC*(M-1))]In which M is the number of switch output ports, d is the link basic delay, C is the link bandwidth, Q PFC And (5) presetting a PFC pause threshold.
Further, the method also comprises the step of judging whether the received packet carries the packet with the ECN mark at the receiving end, if so, copying the ECN mark into the congestion notification packet CNP, periodically sending the congestion notification packet CNP and the ACK signal to the sending end, otherwise, directly sending the ACK signal to the sending end.
Further, the method further comprises the steps of controlling the sending rate of the sending end according to the type of the received data packet, adjusting the sending rate according to an AIMD algorithm when the congestion notification packet CNP (congestion notification packet) is received, and setting the sending rate directly as a target rate when the data packet with the congestion notification message CNM is received and then sending the data packet.
Further, the adjusting the sending rate according to the AIMD (Additive Increase Multiplicative Decrease) additive multiplicative subtractive algorithm comprises:
if the congestion flag in the congestion notification packet CNP is 1, then the following formula is followedCalculating ECN marker proportion->Wherein g is rate adjustment weight, < >>For the ECN mark ratio in the last statistical period, then calculate the new transmission rate after adjustment: />
If the congestion flag in the congestion notification packet CNP is 0, then the following formula is followedCalculating ECN marker proportion->Then calculate the new transmission rate after adjustment: new transmission rate= (target rate+current transmission rate)/2.
Further, when receiving the data packet with the congestion notification message CNM, the sending rate is directly set to the target rate of C/N, where C is the link bandwidth, and N is the number of congestion flows carried by the data packet with the congestion notification message CNM.
A computer readable storage medium storing a computer program which when executed performs a method as described above.
Compared with the prior art, the invention has the advantages that:
1. in the lossless network of the data center, the congestion state of the output port is judged by the switch according to the queue length so as to identify continuous congestion and sudden congestion, and the ECN marking function of the forwarded data packet is started or the congestion notification message CNM is generated according to the congestion state control, so that the congestion signal is directly fed back from the switch to the host end, the rapid congestion sensing can be realized, and the dynamic regulation of the sending rate is carried out according to the congestion state types by marking different congestion states, thereby effectively solving the problems of continuous congestion and sudden congestion.
2. The invention starts the ECN marking function to control and adjust the sending rate according to the end-to-end ECN marking for the continuous congestion state of queue growth caused by long stream competition, which can effectively control the continuous congestion of queue growth caused by long stream competition.
3. The invention marks the congestion flow of the non-congestion flow sharing inlet port by identifying the congestion flow of the non-congestion flow sharing inlet port, generates the congestion notification message CNM and feeds the congestion notification message CNM back to the host end, so that the sudden congestion of the lossless network of the data center can be timely perceived and controlled, and the congestion flow affecting the victim flow can be timely subjected to rate adjustment, thereby avoiding the problems of PFC triggering and the negative effects of PFC such as queue head blocking, and effectively solving the problem that the PFC can be triggered even if an end-to-end transmission protocol is deployed in the lossless network of the data center.
Drawings
Fig. 1 is a schematic implementation flow diagram of a data flow transmission control method based on congestion feedback in a lossless network of a data center of the present embodiment.
Fig. 2 is a schematic diagram of a detailed implementation flow for implementing congestion feedback-based data stream transmission control in a lossless network of a data center in a specific application embodiment of the present invention.
FIG. 3 is a schematic diagram of a test scenario topology of a test bed in a specific application embodiment.
FIG. 4 is a topology diagram of a large-scale simulation test scenario of NS-3 in a specific application embodiment.
FIG. 5 is a schematic diagram of the basic performance test results of the test bed in a specific application example.
FIG. 6 is a graphical representation of performance results under a web search workload in a specific application embodiment.
FIG. 7 is a graphical representation of performance results under data mining workload in a specific application embodiment.
Detailed Description
The invention is further described below in connection with the drawings and the specific preferred embodiments, but the scope of protection of the invention is not limited thereby.
As shown in fig. 1, the steps of the data stream transmission control method based on congestion feedback in the lossless network of the data center of the present embodiment include:
detecting the queue length of the data packet of the output port of the switch (congestion point/notification point), and judging the congestion state of the output port according to the detected queue length;
if the congestion state of the output port is in a congestion-free state, directly forwarding the data packet to the destination output port to be sent to the host end;
if the congestion state of the output port is in a continuous congestion state (continuous congestion of queue growth caused by long-stream competition), starting an ECN marking function, and sending data packets to a host end through a target output port after ECN marking is carried out on specified data packets in a data queue so as to control the sending rate of the host end according to the ECN marking;
if the congestion state of the output port is in a burst congestion state and the data packets of the congestion flow share the input port with the non-congestion flow, the ECN marking function is disabled until the length of the queue drops by a specified threshold value, the data packets are forwarded to the destination output port, and meanwhile, a congestion notification message CNM is generated and the sending end of the congestion flow is directly notified, so that the sending rate of the host end is controlled according to the congestion notification message CNM. The sender of the congestion flow is the host for controlling the rate.
In the lossless network of the data center, the congestion state of the output port is judged by the switch according to the queue length so as to identify continuous congestion and sudden congestion, and the ECN marking function of the forwarded data packet is started or the congestion notification message CNM is generated according to the congestion state control, so that the congestion signal is directly fed back from the switch to the host end to realize quick congestion sensing, and the dynamic adjustment of the sending rate is carried out according to the congestion state types by marking different congestion states, so that the problems of continuous congestion and sudden congestion are effectively solved. For the continuous congestion state, the ECN marking function is started to control and adjust the sending rate according to the end-to-end ECN marking, so that the continuous congestion caused by queue growth due to long-flow competition can be effectively controlled; for burst congestion caused by short flows, by identifying the congestion flow (the flow which really causes congestion) of the shared inlet port of the non-congestion flow (the flow which does not cause network congestion), the congestion flow of the shared inlet port of the non-congestion flow is marked, a congestion notification message CNM is generated and fed back to a host end, so that the burst congestion of a data center lossless network can be timely sensed and controlled, and the rate of the congestion flow which affects the victim flow is timely regulated, thereby avoiding the problems of PFC triggering and negative influences such as queue head blocking, and effectively solving the problem that PFC can be triggered even if an end-to-end transmission protocol is deployed in the data center lossless network.
In this embodiment, determining the congestion state of the output port according to the detected queue length specifically includes: when the length of the queue is smaller than a preset ECN marking threshold QECN, judging that the output port is in a congestion-free state; when the length of the queue is between a preset ECN marking threshold QECN and a preset burst congestion feedback threshold QCNM, the output port is judged to be in a continuous congestion state, namely continuous congestion caused by long-stream competition, and when the length of the queue exceeds the preset burst congestion feedback threshold QCNM, the output port is judged to be in a burst congestion state, namely burst congestion caused by short-stream.
Preferably, the preset ECN marking threshold QECN may specifically be 32KB, and the value of the burst congestion feedback threshold QCNM is within the range Q ECN ，max(Q ECN ,Q PFC /M-3dC*(M-1))]Internal value, wherein M is the number of switch output ports, d is link basic delay, C is link bandwidth, Q PFC And (5) presetting a PFC pause threshold. It can be understood that the preset ECN marking threshold bursty congestion feedback threshold QCNM and the bursty congestion feedback threshold QCNM can be configured according to actual requirements.
In this embodiment, if the congestion state of the output port is in the continuous congestion state, the ECN marking is performed on the data packets in the data queue that exceed the preset ECN marking threshold QECN, so that the sending rate of the corresponding data packets is adjusted at the host end according to the ECN marking, and the continuous congestion that the queue grows due to long-flow contention is effectively controlled.
In this embodiment, if the congestion state of the output port is in the bursty congestion state, it is determined whether the congestion flow to which the data packet belongs shares the input port with the non-congestion flow, if so, the ECN marking function is disabled until the queue length falls below the preset ECN marking threshold QECN, and then the ECN marking function is restarted, otherwise, the data packet is directly forwarded. In the embodiment, by judging whether the congestion flow to which the data packet belongs shares the input port with the non-congestion flow and identifying and marking the congestion flow sharing the input port with the non-congestion flow, only the congestion flow affecting the victim flow is subjected to rate adjustment, the problems of PFC triggering and negative influences of PFC such as queue head blocking can be effectively avoided, and the problem of sudden congestion in a data center lossless network is solved.
The embodiment also comprises the step of judging whether the received packet carries the packet with the ECN mark at the receiving end (the notification point), if so, copying the ECN mark into the congestion notification packet CNP, periodically sending the congestion notification packet CNP and an ACK (acknowledgement) signal to the sending end, and if not, directly sending the ACK signal to the sending end. That is, at the receiving end, after receiving the data packet with the ECN flag, the ECN flag is copied into the congestion notification packet CNP, and the CNP packet is periodically transmitted to the transmitting end.
In this embodiment, the method further includes controlling a sending rate of a sending end (a reaction point) according to a received data packet type, adjusting the sending rate according to an AIMD algorithm when a congestion notification packet CNP is received, and directly setting the sending rate as a target rate when a data packet with a congestion notification packet CNM is received, and retransmitting the data packet. Namely, at a transmitting end, after receiving a congestion notification message CNP packet, adjusting the transmitting rate according to an AIMD algorithm; when the CNM packet is received, the sending rate is directly set as the target rate, and the rate adjustment is carried out on the congestion flow affecting the victim flow in time.
In this embodiment, the adjusting the sending rate according to the AIMD algorithm specifically includes:
if the congestion flag in the congestion notification packet CNP is 1, then the formula is followedCalculating ECN marker proportion->Wherein g is rate adjustment weight, < >>For the ECN mark ratio in the last statistical period (set to one round trip delay RTT), then calculate the adjusted new sending rate: />
If the congestion flag in the congestion notification packet CNP is 0, then the formula is followedCalculating ECN marker proportion->Then calculate the new transmission rate after adjustment: new transmission rate= (target rate+current transmission rate)/2.
When the CNP packet is received, the embodiment determines the adjusted sending rate according to the ECN marking proportion by adopting an AIMD algorithm, and can perform dynamic rate adjustment according to the congestion degree, thereby effectively solving the problem of continuous congestion.
In this embodiment, when a data packet with a congestion notification message CNM is received, the sending rate is specifically set to be the target rate of C/N directly, where C is the link bandwidth, and N is the number of congestion flows carried by the data packet with the congestion notification message CNM. Namely, if the data packet with the congestion notification message CNM is received, the output port is indicated to be in a sudden congestion state currently, the speed reduction processing is needed, and the sending rate suitable for the current congestion state is determined according to the link bandwidth and the congestion flow number, so that the sudden congestion problem is effectively solved.
In a specific application embodiment, as shown in fig. 2, a detailed flow of implementing congestion feedback-based data stream transmission control in a lossless network of a data center among a switch, a receiving end and a transmitting end is as follows:
step 1, exchange side (Congestion Point/Announcement Point)
Step 1.1 at the exchange side, initializing the ECN marking threshold Q ECN For 32KB, the number M of switch output ports is 32, the link basic delay d is 5 mu s, and a burst congestion feedback threshold Q is set CNM And make ECN mark threshold Q CNM The value of [ Q ] ECN ,max(Q ECN ,Q PFC /M-3dC*(M-1))]An inner part;
step 1.2, after receiving the data packet, firstly detecting the congestion state of the port according to the queue length of the output port, judging whether the queue length of the current output port is greater than an ECN mark threshold value, if not, judging that the output port is in a congestion-free state, and directly forwarding the packet to a destination output port; otherwise (the length of the queue is smaller than the ECN marking threshold), judging whether the length of the queue of the current output port is smaller than the burst congestion feedback threshold, if so, starting the ECN marking function and marking ECN bits of the data packets exceeding QECN in the queue; otherwise (the length of the queue exceeds a burst congestion feedback threshold QCNM), judging whether the congestion flow to which the packet belongs shares an input port with the non-congestion flow, if so, disabling an ECN marking function, generating a congestion notification message CNM, carrying the number of the congestion flow of the output port, directly transmitting the CNM to a transmitting end of the congestion flow and forwarding the packet to a destination output port; otherwise (without sharing the ingress port with non-congested flows), the packet is forwarded directly to the destination egress port.
Step 2, receiving end (announcement point)
The receiving end judges whether the current data packet has ECN marks, if so, the ECN marks are copied into congestion notification packets CNP, CNP packets are sent to the sending end periodically (for example, every 55 mu s), and ACK is sent to the sending end; otherwise, the ACK is directly sent to the sending end.
Step 3, sender (reaction Point)
Step 3.1, the sender firstly initializes the link bandwidth C to be set as the bandwidth value of the switch outlet port, the rate adjustment weight g is 0.185, and the congestion flow number N is 0;
step 3.2, judging whether the current data packet is a CNP packet, if so, adjusting the sending rate according to an AIMD algorithm, and then sending the packet, namely if the congestion mark in the CNP is 1, according to the formula:setting a target rate as a current transmission rate, wherein +.>If congestion flag in CNP is 0, then the following formula: (new transmission rate= (target rate+current transmission rate)/2) setting target rate as current transmission rate, whereinOtherwise (not CNP grouping), judging whether the current grouping is a CNM grouping, if so, directly setting the sending rate as a target rate C/N, wherein C is the link bandwidth, N is the number of congestion flow carried by the CNM grouping, and then sending the grouping; otherwise (not CNM packets), the sender sends the packet.
According to the embodiment, through the steps, the continuous congestion and the sudden congestion are distinguished according to the queue length of the output port in the lossless network of the data center, the continuous congestion is solved through the end-to-end congestion feedback, the sudden congestion is solved through the direct congestion feedback of the switch, and only the congestion flow sharing the input port with the non-congestion flow is decelerated, so that the problems of the continuous congestion and the sudden congestion can be effectively solved, and the problems of triggering PFC and blocking of the queue head of the PFC are avoided.
In order to verify the effectiveness of the method, the method is respectively realized in a real network test bed and an NS-3 network simulation platform, and performance tests are carried out. FIG. 3 is a test scenario topology of a test bed. The test bed adopts a typical leaf-spine network topology structure of a data center, and comprises 3 leaf switches (L0, L1 and L2), 2 spine switches (S0 and S1), 16 sending ends (H0-H15) and 2 receiving ends (R0 and R1). In a test bed experiment environment, each host (Dell PRECISION TOWER 5820) is provided with a 10-core Intel Xeon W-2255CPU, a 64GB memory and a DPDK (data plane development kit) Mellanox ConnectX-5 100GbE NIC (network interface controller), the cache size of a programmable P4 (programming protocol-independent packet processors) switch is 22MB, each port enables a PFC mechanism, the PFC PAUSE threshold is 320KB, all link bandwidths are set to 40Gbps, and the basic RTT is 40 microseconds. The transmission control protocol deploys DCQCN with parameters set to default parameters, where ECN marking threshold is 200KB. Two long streams f0 and f1 are transmitted from the transmitting ends H0 and H1 to the receiving ends R0 and R1, respectively. Burst traffic is sent from the sending terminals H2-H15 to the receiving terminal R1, each of which generates 35 short streams of 64KB in size, the burst short streams lasting 8 ms.
NS-3 Large Scale simulation test scenario topology As shown in FIG. 4, a leaf spine network topology is employed with 8 core layer spine switches and 10 Access layer leaf switches. Each leaf switch is connected to all spine switches on the upper layer and simultaneously to 24 terminal servers. The entire leaf spine network topology has 240 servers. All links in the network have a bandwidth of 40Gbps and each hop has a delay of 5 microseconds. Network traffic is generated between randomly selected host pairs and is subject to the poisson arrival process. Experiments select two typical applications for two typical centers of a data center, web search and data mining, both of which are heavy-tailed, in which about 60% of the flow is less than 100KB and about 20% of the flow is greater than 1MB; in a data mining workload, about 80% of the streams are less than 100KB and about less than 5% of the streams are greater than 35MB. The network load gradually increases from 0.4 to 0.8.
The basic performance test result of the test bed is shown in fig. 5, wherein (a) in fig. 5 is a queue length, (b) in fig. 5 is a real-time throughput rate, and DCON is the transmission control method of the present invention. From the figure, the DCON of the invention can timely and accurately realize congestion control, effectively avoid the PFC triggered by the ingress port P1 of the leaf switch L2, and the non-congestion flow f0 does not experience queue head blocking. During the existence of bursty traffic, the DCON congestion awareness and control mechanism effectively controls the rate of the congested flow f1 in time. The queue length of the P1/L2 of the inlet port shared by the non-congestion flow f0 and the congestion flow f1 does not exceed the PFC threshold, so that the P0/S0 port is prevented from being suspended by PFC, and the non-congestion flow f0 is prevented from being blocked. In a word, the DCON of the invention can effectively process sudden congestion, and successfully avoid the problems of congestion and congestion diffusion of the queue head of the PFC.
The performance test results under the web search and data mining workload are shown in fig. 6 and 7, wherein (a) in fig. 6 and (a) in fig. 7 are PFC pause message rates under the web search and data mining workload, respectively, fig. 6 (b) and (b) are average stream completion time charts under the web search and data mining workload, respectively, and fig. 6 (c) and (c) are 99-bit tail stream completion times under the web search and data mining workload, respectively. As can be seen from the figure, the DCON can identify the congestion flow in time, and directly send the congestion notification from the switch to the corresponding sending end, so that the sending rate is directly reduced to the target rate, and the DCON obtains the lowest average flow completion time and 99-bit trailing flow completion time, so that the PFC pause message rate can be effectively reduced. The DCON of the invention effectively avoids the blocking of the queue head of PFC, so that the flow is completed more quickly, and the application performance is effectively improved.
The present embodiment also provides a computer-readable storage medium storing a computer program which, when executed, implements a method as described above.
The foregoing is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. While the invention has been described with reference to preferred embodiments, it is not intended to be limiting. Therefore, any simple modification, equivalent variation and modification of the above embodiments according to the technical substance of the present invention shall fall within the scope of the technical solution of the present invention.
Claims (9)
1. A data stream transmission control method based on congestion feedback in a lossless network of a data center is characterized by comprising the following steps:
detecting the queue length of a data packet of an output port of the switch, and judging the congestion state of the output port according to the detected queue length;
if the congestion state of the output port is in a congestion-free state, directly forwarding the data packet to the destination output port to be sent to the host end;
if the congestion state of the output port is in a continuous congestion state, starting an ECN marking function, and sending data packets to a host end through a target output port after ECN marking is carried out on specified data packets in a data queue so as to control the sending rate of the host end according to the ECN marking;
if the congestion state of the output port is in a burst congestion state and the data packets of the congestion flow share the input port with the non-congestion flow, disabling the ECN marking function until the length of the queue drops by a specified threshold value, forwarding the data packets to the target output port, and generating a congestion notification message CNM and directly notifying a transmitting end of the congestion flow so as to control the transmitting rate of the host end according to the congestion notification message CNM;
the judging the congestion state of the output port according to the detected queue length comprises the following steps: when the length of the queue is smaller than a preset ECN marking threshold QECN, judging that the output port is in a congestion-free state; when the length of the queue is between a preset ECN marking threshold QECN and a preset burst congestion feedback threshold QCNM, judging that the output port is in a continuous congestion state; and when the length of the queue exceeds a preset burst congestion feedback threshold QCNM, judging that the output port is in a burst congestion state.
2. The method according to claim 1, wherein if the congestion state of the output port is in a continuous congestion state, ECN marking is performed on the data packets in the data queue exceeding the preset ECN marking threshold QECN.
3. The method according to claim 1, wherein if the congestion state of the output port is in a bursty congestion state, it is determined whether the input port is shared by the congestion flow to which the data packet belongs and the non-congestion flow, if so, the ECN marking function is disabled until the queue length falls below the preset ECN marking threshold QECN, and if not, the ECN marking function is re-enabled, otherwise, the data packet is forwarded directly.
4. The method for congestion feedback based data streaming control in a data center lossless network according to claim 1, wherein the value of the bursty congestion feedback threshold QCNM is in the range [ Q ] ECN ，max(Q ECN ,Q PFC /M-3dC*(M-1))]In which M is the number of switch output ports, d is the link basic delay, C is the link bandwidth, Q PFC And (5) presetting a PFC pause threshold.
5. The method according to any one of claims 1 to 4, further comprising determining, at the receiving end, whether the received packet carries an ECN-marked packet, if so, copying the ECN-marked packet to a congestion notification packet CNP, periodically transmitting the congestion notification packet CNP and an ACK signal to the transmitting end, and otherwise, directly transmitting the ACK signal to the transmitting end.
6. The method according to claim 5, further comprising controlling a transmission rate of a transmitting end according to a type of a received data packet, adjusting the transmission rate according to an AIMD algorithm when the congestion notification packet CNP is received, and directly setting the transmission rate as a target rate when the data packet with the congestion notification message CNM is received, and retransmitting the data packet.
7. The method for congestion feedback based data streaming control in a data center lossless network according to claim 6, wherein said adjusting the transmission rate according to the AIMD algorithm comprises:
if the congestion flag in the congestion notification packet CNP is 1, then the following formula is followedCalculation of ECN-tag ratioWherein g is rate adjustment weight, < >>For the ECN mark ratio in the last statistical period, then calculate the new transmission rate after adjustment: />
If the congestion flag in the congestion notification packet CNP is 0, then the following formula is followedCalculating ECN marker proportion->Then calculate the new transmission rate after adjustment: new transmission rate= (target rate+current transmission rate)/2.
8. The method according to claim 6, wherein when receiving the data packet with the congestion notification message CNM, the sending rate is directly set to a target rate of C/N, where C is a link bandwidth, and N is a congestion flow number carried by the data packet with the congestion notification message CNM.
9. A computer readable storage medium storing a computer program, which when executed by a processor implements the method of any one of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210676107.2A CN114938350B (en) | 2022-06-15 | 2022-06-15 | Congestion feedback-based data stream transmission control method in lossless network of data center |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210676107.2A CN114938350B (en) | 2022-06-15 | 2022-06-15 | Congestion feedback-based data stream transmission control method in lossless network of data center |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114938350A CN114938350A (en) | 2022-08-23 |
| CN114938350B true CN114938350B (en) | 2023-08-22 |
Family
ID=82865933
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210676107.2A Active CN114938350B (en) | 2022-06-15 | 2022-06-15 | Congestion feedback-based data stream transmission control method in lossless network of data center |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114938350B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115022249B (en) * | 2022-06-17 | 2024-03-22 | 长沙理工大学 | Data stream transmission scheduling control method and system in cloud data center application |
| CN115460156B (en) * | 2022-09-21 | 2024-02-02 | 长沙理工大学 | Method, device, equipment and medium for controlling lossless network congestion of data center |
| CN115396372B (en) * | 2022-10-26 | 2023-02-28 | 阿里云计算有限公司 | Data stream rate control method, intelligent network card, cloud device and storage medium |
| US20240163220A1 (en) * | 2022-11-10 | 2024-05-16 | Huawei Technologies Co., Ltd. | Methods and apparatus for improved congestion signaling |
| CN116489106B (en) * | 2023-06-21 | 2023-09-19 | 新华三技术有限公司 | Congestion control method, device, forwarding chip and client |
| CN116915706B (en) * | 2023-09-13 | 2023-12-26 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | Data center network congestion control method, device, equipment and storage medium |
| CN117395206B (en) * | 2023-12-12 | 2024-03-12 | 南京邮电大学 | Rapid and accurate congestion feedback method for lossless data center network |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108270691A (en) * | 2018-01-11 | 2018-07-10 | 电子科技大学 | A kind of alternative ecn (explicit congestion notification) labeling method for data center |
| CN108718283A (en) * | 2018-05-03 | 2018-10-30 | 电子科技大学 | The TCP jamming control methods that centralized end net is coordinated in data center network |
| CN110061927A (en) * | 2019-04-26 | 2019-07-26 | 东南大学 | Congestion aware and labeling method towards micro- burst flow in a kind of more queuing data center environments |
| CN110177051A (en) * | 2019-06-28 | 2019-08-27 | 四川长虹电器股份有限公司 | Data center's jamming control method based on fluidics |
| WO2020042624A1 (en) * | 2018-08-30 | 2020-03-05 | 华为技术有限公司 | Transmission rate control method and apparatus, sending device and receiving device |
| CN111464452A (en) * | 2020-03-30 | 2020-07-28 | 南京理工大学 | Fast congestion feedback method based on DCTCP |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080298248A1 (en) * | 2007-05-28 | 2008-12-04 | Guenter Roeck | Method and Apparatus For Computer Network Bandwidth Control and Congestion Management |
| US8619573B2 (en) * | 2010-03-24 | 2013-12-31 | Telefonaktiebolaget Lm Ericsson (Publ) | Delayed flow control action in transport network layer WCDMA communications |
| CN109218222B (en) * | 2017-06-30 | 2021-06-22 | 华为技术有限公司 | Method, device and system for realizing speed regulation of sending end |
| CN108540395B (en) * | 2018-04-18 | 2020-07-28 | 清华大学 | Congestion judgment method in loss-free network and switch |
-
2022
- 2022-06-15 CN CN202210676107.2A patent/CN114938350B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108270691A (en) * | 2018-01-11 | 2018-07-10 | 电子科技大学 | A kind of alternative ecn (explicit congestion notification) labeling method for data center |
| CN108718283A (en) * | 2018-05-03 | 2018-10-30 | 电子科技大学 | The TCP jamming control methods that centralized end net is coordinated in data center network |
| WO2020042624A1 (en) * | 2018-08-30 | 2020-03-05 | 华为技术有限公司 | Transmission rate control method and apparatus, sending device and receiving device |
| CN110061927A (en) * | 2019-04-26 | 2019-07-26 | 东南大学 | Congestion aware and labeling method towards micro- burst flow in a kind of more queuing data center environments |
| CN110177051A (en) * | 2019-06-28 | 2019-08-27 | 四川长虹电器股份有限公司 | Data center's jamming control method based on fluidics |
| CN111464452A (en) * | 2020-03-30 | 2020-07-28 | 南京理工大学 | Fast congestion feedback method based on DCTCP |
Non-Patent Citations (1)
| Title |
|---|
| Mahmoud Bahnasy ; Halima Elbiaze.Fair Congestion Control Protocol for Data Center Bridging.IEEE Systems Journal ( Volume: 13, Issue: 4, December 2019).2019,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114938350A (en) | 2022-08-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114938350B (en) | Congestion feedback-based data stream transmission control method in lossless network of data center | |
| EP0688481B1 (en) | Device and method for atm end system cell flow regulation | |
| CN111316605B (en) | Layer 3 fair rate congestion control notification | |
| CN113507420B (en) | Congestion control method and device | |
| US7859996B2 (en) | Intelligent congestion feedback apparatus and method | |
| US6535482B1 (en) | Congestion notification from router | |
| CN109714267B (en) | Transmission control method and system for managing reverse queue | |
| EP3107252B1 (en) | Fast friendly start for a data flow | |
| CN112104562B (en) | Congestion control method and device, communication network and computer storage medium | |
| CA2237264A1 (en) | Receiver based congestion control | |
| CN112995048B (en) | Blocking control and scheduling fusion method of data center network and terminal equipment | |
| Sisalem et al. | LDA+: A TCP-friendly adaptation scheme for multimedia communication | |
| WO2021036534A1 (en) | Link bandwidth adjustment method and apparatus, and storage medium | |
| CN115134302B (en) | Traffic isolation method for avoiding queue head blocking and congestion diffusion in lossless network | |
| CN113783785A (en) | ECN (engineering-centric networking) water line value configuration method and device and network equipment | |
| CN115134308A (en) | Method for avoiding head of line blocking through data packet bouncing in lossless network of data center | |
| CN114785743A (en) | Active queue management algorithm for enhancing RED stability | |
| US11622028B2 (en) | Explicit notification of operative conditions along a network path | |
| CN115883463A (en) | Network load-based congestion control method and system in data center | |
| JP2008193310A (en) | Buffer management method of router and router using the management method | |
| CN115914106B (en) | Self-adaptive buffer method for network traffic forwarding | |
| JP2006087010A (en) | Communication path control apparatus and network system employing the same | |
| Semeria et al. | Supporting Differentiated Service Classes: Active Queue Memory Management | |
| Kim et al. | TCP acknowledgement compression for fairness among uplink TCP flows in IEEE 802.11 n WLANs | |
| CN114285803A (en) | Congestion control method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |