JP5324637B2 - Dynamic flowlet scheduling system, flow scheduling method, and flow scheduling program - Google Patents

Abstract

The invention provides a dynamic flow division scheduling system and a dynamic flow division scheduling method for a data center network. A host machine applies for bandwidth reservation of specific flow for a flow scheduler of a data center; the flow scheduler checks the bandwidth using conditions of all paths which can be selected by the specific flow; if the sum of bandwidths retained on the paths meets the demand for the bandwidth of the specific flow, the flow is divided into one or more sub-flows, and then the sub-flows are forwarded on the corresponding paths; and if the bandwidths retained on all the paths cannot meet the demand for the bandwidth of the specific flow, the current flows on the paths are transferred to other paths, so that a space is reserved for forwarding the specific flow. According to the dynamic flow division scheduling system and the dynamic flow division scheduling method, the actual bandwidth demand of the flow can be timely and accurately obtained, and scheduling can be performed on the granularity which is thinner than the flow, so that the utilization rate of the whole data center network is balanced, and the convergence bandwidth of the whole data center network is increased.

Description

  The present invention relates to the field of data center networks, and more particularly to systems and methods for dynamic flowlet scheduling.

  With the development of applications such as the Internet and cloud computing services, the scale of data centers is increasing, and it is not uncommon to now have tens of thousands to hundreds of thousands of servers and switches. In addition, data centers typically support a wide variety of applications simultaneously, including those requiring large amounts of data communication between different servers in the data center. These data center expansions and application developments are creating new challenges for data center network architectures.

  In general, since the number of ports of a switch is limited, the data center has a tree topology format with multiple roots. FIG. 1 shows a multi-root tree topology that provides a plurality of arbitrary paths for inter-host communication. Since there are multiple arbitrary paths in this topology, it is an important issue to establish a method for dynamically scheduling flows between these paths in order to achieve highly aggregated bandwidth.

  Currently, the most commonly used mechanism in flow scheduling is the equivalent cost multipath (ECMP) scheme. Given a single destination host, the switch is configured such that multiple forwarding arbitrary paths are established for that host. When a packet arrives at the switch, the switch performs a hash operation on a particular field in the packet header, followed by a modulo operation on the total number of arbitrary paths to determine the final selected path, Forward the packet through the path. The hash algorithm ensures a load balance by randomly distributing data traffic among a plurality of arbitrary paths.

  However, since ECMP is a static mapping scheme, all packets of the same flow are always mapped to one specific path. Further, according to ECMP, the network usage status and flow size are ignored, and a transfer path is selected based only on the result of the hash calculation. With such a method, the same path may be selected for a plurality of large-scale flows depending on the situation. When this occurs, a bottleneck occurs in the network and the usage rate of the entire network decreases. FIG. 2 shows two types of flow collisions. Referring to FIG. 2, since a hash collision occurs in the switch Agg0, the same path is selected for the flow A and the flow B. In addition, packets of flow C and flow D transferred separately from the switches Agg1 and Agg2 collide at the switch Core2. If the maximum bandwidth of each link in FIG. 2 is 1 Gbps, all four flows can only reach a bandwidth of 500 Mbps due to collisions. If there is a method superior to ECM that can schedule flows based on network resource usage and flow states, for example, flow A is scheduled on switch Core1 and flow D is scheduled on switch Core3. Each of the flows can reach a bandwidth of 1 Gbps.

  As described above, ECMP, which is a static mapping scheme, does not consider both the network usage status and the flow size, so that a collision occurs and the overall network usage rate decreases.

  Non-Patent Document 1 ("Hedera: Dynamic Flow Scheduling for Data Center Networks (Hedera: Dynamic Flow Scheduling for Data Center Networks)", Mohammad Al-Fares, et al, NSDI 2010) A dynamic flow scheduling system is proposed. First, Hedera detects a large number of flows with an edge switch. Next, it anticipates bandwidth demands due to large flows and calculates the best path for these flows. Finally, the calculated path is installed on the switch.

FIG. 3 is a structural diagram of the Hedera system. As shown in FIG. 3, Hedera is a flow scheduling system designed and implemented based on a “Fat Tree” network topology (Non-Patent Document 2 (“A scalable, Community Data Center Network Architecture (Scalable Commodity Data Center Network Architecture ) ", Mohammad Al-Fares, et al, SIGCOMM 2008)). The switches in the fat tree are organized into edge layers, aggregation layers, and core layers (from bottom to top). All switches in the three different layers are the same, each with k ports. FIG. 3 shows a fat tree topology with k = 4. The fat tree is divided into k pods, which are indicated by k dotted frames in the edge layer and the aggregation layer. k switches surrounded by each of the k dotted line frames constitute one pod. The fat tree topology, each by 5k ^2/4 pieces of switch has k port, can support k ^3/4 pieces of the host. In each switch, OpenFlow is operating. The OpenFlow controller can access the forwarding table in each switch and execute operations such as query, insertion, and correction on the flow entry in the forwarding table. The flow scheduler schedules a flow in a path where no collision occurs based on the current network state. The OpenFlow controller and the flow scheduler run on a logically centralized central management device.

  FIG. 4 is a block diagram of the Hedera system 40. As shown in FIG. 4, the Hedera system 40 includes a switch 410 and a central management device 420.

  The switch 410 includes a forwarding table 4110, a packet forwarding unit 4120, and a flow statistics reporting unit 4130. The forwarding table 4110 holds mapping between flows and transmission ports. The packet transfer unit 4120 transfers the packet based on the mapping held in the transfer table 4110. The flow statistics reporting means 4130 generates traffic statistics and reports the statistics to the central management apparatus 420 periodically (eg, every 5 seconds).

  The central management apparatus 420 includes a switch controller 4210, a flow scheduling unit 4220, a flow request prediction unit 4230, a flow request table 4240, and a flow statistics collection unit 4250. The flow statistics collection unit 4250 receives traffic statistics from the flow statistics report unit 4130 during each scheduling period (eg, 5 seconds), and detects a large number of flows. The flow request table 4240 holds a flow request matrix for all hosts. The flow request prediction means 4230 uses the traffic statistics to predict a flow bandwidth request. The flow scheduling means 4220 selects and assigns an appropriate path from all arbitrary paths to a large quantity of flows. The switch controller 4210 creates a forwarding entry in the corresponding switch based on the scheduling result received from the flow scheduling unit 4220 so that a large amount of flow is forwarded to the newly selected path.

  Specifically, the default behavior that switch 410 takes when a new flow starts is to select one of the equivalent cost paths based on the hash value of the flow (explained in connection with ECMP). is there. Switch 410 forwards the flow along this path and continues until the traffic volume of the flow exceeds a predetermined threshold. As the threshold value, for example, 100 Mbps (10% of the bandwidth of the 1 GigE link) can be considered. The flow statistics collection means 4250 obtains traffic statistics once from the edge switch in order to detect a large number of flows during each scheduling period. Thereafter, the flow request prediction unit 4230 predicts a bandwidth request for a large amount of flows. Next, the flow scheduling means 4220 calculates an appropriate path for the mass flow based on the bandwidth request of the mass flow. Finally, the switch controller 4210 installs a path on the corresponding switch by modifying the forwarding entry so that the mass flow is forwarded to this newly selected path.

  Current traffic statistics for all large flows are input to a flow request predictor 4230 that holds an N × N matrix M, where N is the number of hosts. An element M (i, j) in the i-th row and j-th column of the matrix M holds a flow prediction request from the host i to the host j. The flow request prediction means 4230 repeatedly executes the operation of increasing the bandwidth request of the flow at the source host and decreasing the flow bandwidth by an excess amount based on the bandwidth limit of the destination host. This continues until the width requirement is met.

  After a bandwidth request for a large flow is expected, the flow scheduling means 4220 calculates an appropriate path for the large flow based on the bandwidth request. In Non-Patent Document 1, two scheduling algorithms are used. The first scheduling algorithm is called “Global First Fit”. When a new mass flow is detected, the scheduler searches all arbitrary paths linearly and can accommodate all the bandwidth requests of the flow. And find the path as a scheduling result. The second scheduling algorithm is called “Simulated Annealing”, and is a general-purpose stochastic algorithm that searches the state space to find a nearly optimal solution. In order to reduce the number of states in the search space, the Hedera system 40 does not assign one core switch to each flow, but assigns one core switch to each destination host.

“Hedera: Dynamic Flow Scheduling for Data Center Networks (Hedera: Dynamic Flow Scheduling for Data Center Networks)”, Mohammad Al-Fares, et al, NSDI 2010. “A scalable, Community Data Center Network Architecture (Scalable Commodity Data Center Network Architecture)”, Mohammad Al-Fares, et al, SIGCOMM 2008

  However, the scheduling mechanism used in the Hedera system has the following drawbacks. First, the actual bandwidth demand of the flow is expected from the traffic statistics measured at the edge switch, but this statistic sometimes does not accurately reflect the actual bandwidth demand of the application, so Bandwidth requests cannot be obtained accurately and in a timely manner. In addition, since scheduling is performed only once in each scheduling period (eg, 5 seconds), there is always a possibility that the traffic state has fluctuated significantly during one scheduling period. Second, scheduling based on flow granularity is insufficient in some situations to achieve balanced network usage and highly aggregated network bandwidth.

  To address the above problems, the present invention provides a dynamic flowlet scheduling system for a data center network. The host applies to the data center flow scheduler for the bandwidth reserved for the specific flow. The flow scheduler checks the usage status of the bandwidths of all paths through which the specific flow can pass. If the sum of the remaining bandwidths of these paths satisfies the bandwidth requirement of the specific flow, the flow is divided into one or more flowlets and then transferred to the corresponding path. If the total remaining bandwidth of these paths does not meet the bandwidth requirements for that particular flow, the existing flows on these paths will have other flows to make room for forwarding that particular flow. Moved to the path.

  According to the first aspect of the present invention, a bandwidth management unit is a central management device configured to collect bandwidth requests of a specific flow from a transmission source host, and bandwidth usage status of a link. And one or more paths from the arbitrary path of the specific flow are allocated for the specific flow based on the bandwidth usage status of the current link, and storage means configured to store the arbitrary path of the host pair, A flow scheduling unit configured to return the assigned path to the transmission source host and receive flowlet information from the transmission source host, and to set a corresponding switch based on a scheduling result of the flow scheduling unit A central management device is provided comprising a configured switch controller.

  The flow scheduling means calculates the total idle bandwidth in an arbitrary path of the specific flow, and if the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, It is desirable to configure one or more paths from an arbitrary path of a specific flow for the specific flow.

  The flow scheduling means further rearranges the arbitrary paths of the specific flow in descending order of the idle bandwidth when the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, It is desirable that the paths are sequentially selected from the rearranged arbitrary paths until the total idle bandwidth of the selected path is equal to or greater than the bandwidth request of the specific flow.

  The flow scheduling means further detects another flow passing through the arbitrary path of the specific flow when the total idle bandwidth in the arbitrary path of the specific flow is lower than the bandwidth request of the specific flow, and detects the other flow Calculate the total idle bandwidth in any path of the flow, and reschedule the other flow if the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request for the specific flow Thus, it is desirable that one or more paths are allocated for the specific flow from the arbitrary paths of the specific flow.

  Furthermore, the flow scheduling means further increases the idle bandwidth total in the arbitrary path when the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request of the specific flow. Select one flow from the rearranged flows in order, and select until the total idle bandwidth in any path of the specific flow is equal to or greater than the bandwidth request of the specific flow It is desirable to be configured to reschedule the flow.

  The flow scheduling means further re-schedules the other flow to the idle path when the total idle bandwidth in the arbitrary path of the other flow is lower than the bandwidth request of the specific flow. It is desirable to allocate the idle bandwidth in the arbitrary path to the specific flow and the remaining flow that still passes through the arbitrary path of the specific flow even after rescheduling based on the Maximin fairness principle.

  The flow scheduling means is preferably further configured to transmit the flowlet information received from the source host to the destination host.

  The storage means is preferably further configured to store the flowlet information.

  The flowlet information preferably includes the port number of the flowlet.

  According to another aspect of the present invention, there is provided a flow scheduling method for collecting a bandwidth request for a specific flow from a source host, and based on a current link bandwidth usage state, Assigning one or more paths from an arbitrary path for the specific flow, returning the assigned path to the transmission source host, receiving flowlet information from the transmission source host, and responding based on the scheduling result And configuring a switch to perform.

  The allocation step includes a step of calculating a total idle bandwidth in an arbitrary path of the specific flow, and a case where the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow. Preferably includes a step of allocating one or more paths from an arbitrary path of the specific flow for the specific flow.

  If the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, the arbitrary path of the specific flow is rearranged in descending order of idle bandwidth. Preferably, one path is selected from the selected arbitrary paths, and this is continued until the total idle bandwidth of the selected path is equal to or greater than the bandwidth request of the particular flow.

  When the total idle bandwidth in the arbitrary path of the specific flow is less than the bandwidth request of the specific flow, a flow passing through the arbitrary path of the specific flow is detected, and the idle in the arbitrary path of the other flow is detected. If the total bandwidth is calculated and the total idle bandwidth in any path of the other flow is equal to or greater than the bandwidth request for the specific flow, the other flow is rescheduled and the It is desirable that one or more paths from an arbitrary path of a specific flow are assigned for the specific flow.

  If the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request of the specific flow, the other flows are rearranged in descending order of the total idle bandwidth in the arbitrary path. Until one flow is selected in order from the rearranged flows, the flow is rescheduled, and the total idle bandwidth in any path of the specific flow is equal to or greater than the bandwidth request of the specific flow This is preferably continued.

  If the total idle bandwidth in the arbitrary path of the other flow is less than the bandwidth request of the specific flow, the other flow is rescheduled to the idle path, and the idle bandwidth in the arbitrary path of the specific flow is Preferably, the width is assigned to the specific flow and the residual flow that still passes through the arbitrary path of the specific flow after rescheduling based on the Maximin fairness principle.

  The method preferably further comprises the step of transmitting the flowlet information received from the source host to the destination host.

  The flowlet information preferably includes the port number of the flowlet.

  According to another aspect of the present invention, the host is a bandwidth reservation unit configured to apply a bandwidth for a specific flow to the central management device, and based on a scheduling result returned from the central management device. To divide the specific flow into one or more flowlets and establish a corresponding connection to transfer one or more flowlets on the path assigned by the central management unit A host comprising a configured flow control means is provided.

  The host preferably further comprises data reconstruction means configured to reconstruct data received via one or more flowlets.

  The flow control means is preferably configured to adjust the corresponding connection when a command to change the flowlet is received from the central management device.

  By using the dynamic flowlet scheduling system and dynamic flowlet scheduling method of the present invention, the actual bandwidth requirements of the flow can be obtained accurately and timely, and the scheduling is based on a finer granularity than the flow. (Ie, based on the granularity of the flowlet). Therefore, the dynamic flowlet scheduling system and the dynamic flowlet scheduling method of the present invention can balance the network usage rate of the entire data center network and further increase the aggregate bandwidth of the entire data center network. It becomes.

  These and other features of the present invention will become more apparent when taken in conjunction with the detailed description and accompanying drawings.

It is the schematic which shows the tree topology of the multiple root | route in a prior art. It is the schematic which shows two types of flow collision in a prior art. It is the schematic which shows the Hedera system in a prior art. It is a block diagram which shows the Hedera system in a prior art. 1 is a block diagram illustrating a dynamic flowlet scheduling system according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating an operation of a dynamic flowlet scheduling system according to an exemplary embodiment of the present invention. 4 is a flowchart illustrating a flow scheduling algorithm according to an embodiment of the present invention. FIG. 6 is a block diagram illustrating a dynamic flowlet scheduling system according to another embodiment of the present invention. 6 is a flowchart illustrating an operation of a dynamic flowlet scheduling system according to another embodiment of the present invention.

  In the following, specific embodiments according to the present invention will be described with reference to the accompanying drawings so that the principle and implementation of the present invention will become more apparent. It should be noted that the present invention is not limited to the specific embodiments described below. In order to avoid complications, detailed descriptions of well-known techniques not directly related to the present invention are omitted.

  FIG. 5 is a block diagram illustrating a dynamic flowlet scheduling system 50 according to one embodiment of the present invention. As shown in FIG. 5, the dynamic flowlet scheduling system 50 includes three types of nodes: a host 510, a central management device 520, and a switch 530.

  The switch 530 includes a transfer table 5310 and a packet transfer unit 5320. The forwarding table 5310 holds mapping between flows and transmission ports. The packet transfer unit 5320 transfers the packet based on the mapping held in the transfer table 5310.

  The host 510 includes a flow control unit 5110 and a bandwidth reservation unit 5120. The bandwidth reservation unit 5120 applies to the central management device 520 for the bandwidth reserved for the specific flow. The flow control unit 5110 divides the specific flow into one or more flowlets based on the scheduling result from the central management apparatus 520, and transfers the flowlets on each path.

  The central management apparatus 520 includes a switch controller 5210, a flow scheduling unit 5220, a bandwidth usage / path information table storage unit 5230, and a bandwidth request collection unit 5240. The bandwidth request collection unit 5240 collects a flow bandwidth request from the host 510. The bandwidth usage / path information table storage unit 5230 stores the bandwidth usage of all links and all arbitrary paths of each host pair. The flow scheduling means 5220 allocates one or more appropriate paths from all arbitrary paths for a specific flow based on the current bandwidth usage of the network link. In order to transfer the specific flow, the switch controller 5210 creates a transfer entry in the corresponding switch based on the scheduling result received from the flow scheduling unit 5220.

  Next, the operation of the dynamic flowlet scheduling system 50 shown in FIG. 5 will be described in detail with reference to FIG.

  FIG. 6 is a flowchart illustrating the operation of the dynamic flowlet scheduling system 50 according to an embodiment of the present invention. In FIG. 6, each of the host A, host B, and host C is implemented using the host 510 shown in FIG. 5, the central management apparatus is implemented using the central management apparatus 520 shown in FIG. It can be implemented using the switch 530 shown in FIG. The specific operation is as follows.

1.1: On the transmission source host A, a specific application notifies the bandwidth reservation unit of a bandwidth request for a flow corresponding to this application.
1.2: The bandwidth reservation means of the transmission source host A applies for bandwidth reservation to the central management apparatus.
1.3: The flow scheduling unit of the central management apparatus executes flow scheduling based on a scheduling algorithm (described later) (that is, one or more paths from all arbitrary paths are allocated for the specific flow). .
1.4: The central management apparatus returns the allocated path and the reserved bandwidth of each path to the transmission source host A that has applied for bandwidth reservation.
1.5: The flow control means of the source host A divides the specific flow into one or more flowlets based on the scheduling result.
1.6: The transmission source host A reports the flowlet information (port number used by the flowlet, etc.) to the central management apparatus.
1.7: The switch controller of the central management device creates a forwarding entry in the corresponding switch so that the corresponding switch forwards the flowlet through the newly selected path.
1.8: Depending on the situation, the central management unit may allow some existing flows to other hosts (eg host C) to make room for the bandwidth reservation of the particular flow. Instruct to change path and occupied bandwidth.
1.9: The host C that has received this command changes the paths and occupied bandwidths of the several existing flows in accordance with instructions from the central management apparatus.
1.10: Source host A sends data over one or more connections each corresponding to one flowlet.
1.11: The destination host B receives data via the one or more flowlets.

  FIG. 7 is a flowchart illustrating a flow scheduling algorithm according to an embodiment of the present invention. This scheduling algorithm may be executed by the flow scheduling means 5220 of the central management device 520 shown in FIG. 5 in order to execute flow scheduling.

First, in step S701, the flow F and its bandwidth request _RF are acquired.

  In step S702, all arbitrary paths of the flow F are detected. Here, it is assumed that there are a total of N paths {Pi, 1 <= i <= N} each having an idle bandwidth Ci.

In step S703, the total idle state bandwidth of all N paths is calculated using the following equation:

Next, in step S704, _{S F} is determined whether or _{R F.} If S _F is greater than R _F (This means that that the idle bandwidth of the existing of the N paths meet the bandwidth requirements of the flow F), the process proceeds to step S705. If _SF is less than R _F (this means that the idle bandwidth of the existing N paths does not meet the bandwidth requirements of flow F, so other flows in these paths will make room for flow F to flow The process proceeds to step S707.

が満足されるまでこれが繰り返される。次にステップＳ７０６において、選択されたＭ個のパスがフローＦに割り当てられ、Ｍ個のパスのアイドル帯域幅ＣｉがフローＦのために予約される。これでフロースケジューリング処理は終了し、制御が返される。 If S _F is greater than R _F is, in step S705, the arbitrary path Pi of the specific flow is sorted in descending order of the idle bandwidth Ci, some of these rearranged any path path Pi ( This number is assumed to be M) and the condition

This is repeated until is satisfied. Next, in step S706, the selected M paths are assigned to flow F and the idle bandwidth Ci of M paths is reserved for flow F. This ends the flow scheduling process and returns control.

If S _F is less than R _F , in step S707, another flow {Fj, 1 <= j <= L} that passes through {Pi, 1 <= i <= N} (that is, the same path as the flow F) Are detected). The idle bandwidth of the existing N paths cannot satisfy the bandwidth requirement of the flow F, so some of the L flows need to make room to satisfy the bandwidth requirement of the flow F. There is.

に共通の要素を有することができる。 In step S708, all arbitrary paths {Hjk, 1 <= k <= Nj} are detected for each flow Fj. Here, the j-th flow Fj has Nj paths, and each path Hjk has an idle bandwidth Cjk. Further, in the path {Hjk, 1 <= k <= Nj, 1 <= j <= L}, it is assumed that there are a total of T paths, and each of the T paths has an idle bandwidth Ct. . A plurality of different flows Fj are represented by respective path sets Hjk,

Can have common elements.

と、各フローＦｊのアイドル帯域幅の合計

とが、計算される。 In step S709, the total idle bandwidth of the T paths

And the total idle bandwidth of each flow Fj

And is calculated.

In step S710, the sum Sum of the idle bandwidth of the T path is determined whether or _{R F.} If Sum is greater than or equal to R _F , the L flows Fj are rescheduled to another path to make room for the flow F to flow, and the bandwidth requirement of the flow F is satisfied, so the process is step S711. Proceed to If Sum is less than R _F , the flow proceeds to step S712 because the bandwidth request for flow F is not satisfied even if L flows Fj are rescheduled to another path.

If Sum is greater than or equal to R _F , in step S711, the flow Fj is rearranged in order of increasing total idle bandwidth of the path Sj, several flows Fj are selected in turn, and these flows are assigned to their respective idle paths. And the freed bandwidth is allocated to flow F. This operation, S _F is repeatedly continued until the above R _F. The rescheduling operation is composed of subdivision of existing flowlets, aggregation, and bandwidth change. When S _F is greater than or equal to _{R F,} the process proceeds to step S705.

If Sum is less than R _F , there is no possibility that there is no idle bandwidth in any path of flow F due to the addition of flow F in step S712, but other paths in Fj may have idle bandwidth Sj. Considering that, Fj is rescheduled to each idle path and its bandwidth is exhausted. Thereafter, in step S713, considering that the bandwidth requirement of the flow F is not satisfied even if the L flows Fj are rescheduled to another path, the idle of the path {Pi, 1 <= i <= N} Bandwidth is allocated to flow F and the residual flow that still passes through any of the paths {Pi, 1 <= i <= N} after rescheduling based on the Maximin fairness principle. At this time, the path Pi and the calculated corresponding bandwidth are allocated to the flow F. This ends the flow scheduling process and returns control.

  After the flow scheduling is completed, the flow control unit 5110 of the host 510 acquires a scheduling result from the central management apparatus 520. This result includes the path assigned to the specific flow and the reserved bandwidth of each path. Based on the assigned path, the flow control means 5110 divides the specific flow into one or more flowlets each corresponding to one assigned path, and then one corresponding to each one flowlet. The above connection is established. The flow control means 5110 thus transmits individual data traffic based on bandwidth reservations on different paths.

  When the flow control unit 5110 receives a command for changing the flow corresponding to a certain connection from the central management apparatus 520 during data transmission, the flow control unit 5110 can terminate the connection and start a new connection. In the corresponding flow change, operations such as flow splitting, flowlet aggregation, and flowlet reserved bandwidth change are executed.

  In addition, the flow control means 5110 controls data traffic of higher layer applications during data transmission. When the upper layer application transmits data traffic at a higher speed than the reserved bandwidth, the flow control means 5110 caches the data of the upper layer application and based on the bandwidth reserved by the central management device 520. The data can be transmitted. If the data transmission rate does not reach the reserved bandwidth even after a certain period of time because a large bandwidth is reserved for the application, the flow control means 5110 notifies the central management device 520 , You can ask to change the resource reservation.

  When the application ends, the flow control means 5110 can notify the central management device 520 and request to cancel the reservation of the corresponding resource.

The flow scheduling means 5220 of the central management device 520 needs to know the bandwidth usage status of all links and all arbitrary paths of each host pair during flow scheduling. For this reason, the bandwidth usage status / path information table storage means 5230 includes a path table that holds all arbitrary paths of each host pair, a flow to which each flow belongs, a path through which each passes, and A bandwidth usage status table holding the reserved bandwidths of these paths is stored. Tables 1 and 2 below show examples of a path table and a bandwidth usage table, respectively.

  Some applications, such as media streaming, need to keep received data in that order. However, since the data is divided into a plurality of flowlets at the host and then transmitted through a plurality of connections, the order of the data is not guaranteed. Therefore, the host on the receiving side needs to perform data reconstruction.

  FIG. 8 is a block diagram illustrating a dynamic flowlet scheduling system according to another embodiment of the present invention that allows reconstruction of received data. The dynamic flowlet scheduling system 80 shown in FIG. 8 is different from the dynamic flowlet scheduling system 50 shown in FIG. 5 in that the host 810 of the dynamic flowlet scheduling system 80 further includes data restructuring means 8130. The bandwidth usage status / path information table storage unit 8230 of the central management device 820 stores different contents from the bandwidth usage status / path information table storage unit 5230. Specifically, in this embodiment, the data restructuring unit 8130 constructs data received from a plurality of flowlets, and the bandwidth usage status / path information table storage unit 8230 includes the bandwidth usage status of all links, In addition, not only all arbitrary paths of each host pair but also flowlet information such as a port number is stored. Next, the operation of the dynamic flowlet scheduling system 80 shown in FIG. 8 will be described in detail with reference to FIG.

  FIG. 9 is a flowchart illustrating the operation of the dynamic flowlet scheduling system 80 according to an embodiment of the present invention. In FIG. 9, each of host A, host B, and host C is implemented using the host 810 shown in FIG. 8, the central management device is implemented using the central management device 820 shown in FIG. It can be implemented using the switch 830 shown in FIG.

  As can be seen from FIG. 9, the operations 1.1 to 1.11 in FIG. 9 are the same as those in FIG. 6, and thus detailed description of the operations 1.1 to 1.11. Compared to FIG. 6, FIG. 9 further comprises two operations, 1.12 and 1.13.

  In order for the destination host B to reconstruct the data received from multiple flowlets so that the host application can receive the data in the original order, the data restructuring means determines which flowlet belongs to the same flow. It is necessary to know. Therefore, the central management apparatus provides the flowlet information to the destination host B (operation 1.12.). Thereafter, the destination host B receives data from a plurality of flowlets, and reconstructs the received data based on the flowlet information provided from the central management apparatus (operation 1.13).

  Thus, even if the data is split into multiple flowlets and sent over multiple connections, the dynamic flowlet scheduling system 80 restores the order of the received data and allows for specific applications such as media streaming. Can meet the requirements.

  The present invention has been described above with reference to preferred embodiments. However, the present invention is not necessarily limited to the above embodiments, and various modifications can be made within the scope of the technical idea. it can.

  Further, a part or all of the above-described embodiment can be described as in the following supplementary notes, but is not limited thereto.

(Appendix 1)
A bandwidth request collection means configured to collect bandwidth requests for a specific flow from a source host;
Storage means configured to store link bandwidth usage and an arbitrary path of the host pair;
Based on the bandwidth usage status of the current link, one or more paths from the arbitrary path of the specific flow are allocated for the specific flow, the allocated path is returned to the transmission source host, and the flowlet is transmitted from the transmission source host. Flow scheduling means configured to receive information;
A central controller comprising: a switch controller configured to set a corresponding switch based on a scheduling result of the flow scheduling means.

(Appendix 2)
The flow scheduling means calculates a total idle bandwidth in an arbitrary path of the specific flow, and when the total idle bandwidth in an arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, The central management device according to appendix 1, wherein one or more paths from an arbitrary path of the specific flow are allocated for the specific flow.

(Appendix 3)
The flow scheduling means rearranges the arbitrary paths of the specific flow in descending order of the idle bandwidth when the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, One path is sequentially selected from the rearranged arbitrary paths until the total idle bandwidth of the selected path is equal to or greater than the bandwidth request of the specific flow. The central management device according to attachment 2.

(Appendix 4)
The flow scheduling means detects another flow passing through the arbitrary path of the specific flow when the total idle bandwidth in the arbitrary path of the specific flow is lower than the bandwidth request of the specific flow, and detects the other flow If the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request for the specific flow, the other flow is rescheduled. The central management device according to appendix 2, wherein one or more paths from an arbitrary path of the specific flow are allocated for the specific flow.

(Appendix 5)
If the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request of the specific flow, the flow scheduling means determines that the other flow has a large total idle bandwidth in the arbitrary path. Reorder in order, select one flow from the reordered flows in order, and the selected flow until the total idle bandwidth in any path of the specific flow is equal to or greater than the bandwidth request of the specific flow The central management device according to appendix 4, wherein the central management device is configured to reschedule.

(Appendix 6)
The flow scheduling means reschedules the other flow to the idle path when the total idle bandwidth in the arbitrary path of the other flow is lower than the bandwidth request of the specific flow, and determines the arbitrary path of the specific flow. The additional bandwidth is configured to allocate the idle bandwidth in the specific flow and a remaining flow that still passes through an arbitrary path of the specific flow even after rescheduling based on the Maximin fairness principle Central management device.

(Appendix 7)
The central management device according to claim 1, wherein the flow scheduling means is configured to transmit the flowlet information received from the transmission source host to a destination host.

(Appendix 8)
The central management device according to appendix 7, wherein the storage means is further configured to store flowlet information.

(Appendix 9)
9. The central management device according to appendix 7 or appendix 8, wherein the flowlet information includes a port number of the flowlet.

(Appendix 10)
Collecting bandwidth requests for a specific flow from the source host;
Assigning one or more paths from any path of the specific flow for the specific flow based on bandwidth usage of the current link;
Returning the assigned path to the source host; receiving flowlet information from the source host;
And setting a corresponding switch based on a scheduling result.

(Appendix 11)
Said assigning step comprises:
Calculate the total idle bandwidth in any path of the specific flow;
When the total idle bandwidth in an arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, one or more paths from the arbitrary path of the specific flow are allocated for the specific flow. The flow scheduling method according to Supplementary Note 10.

(Appendix 12)
Said assigning step comprises:
When the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, the arbitrary path of the specific flow is rearranged in descending order of the idle bandwidth, and the idle bandwidth of the selected path 12. The flow scheduling method according to appendix 11, wherein one path is sequentially selected from the rearranged arbitrary paths until the total width becomes equal to or greater than the bandwidth request of the specific flow.

(Appendix 13)
Said assigning step comprises:
When the total idle bandwidth in the arbitrary path of the specific flow is lower than the bandwidth request of the specific flow, another flow passing through the arbitrary path of the specific flow is detected, and the idle bandwidth in the arbitrary path of the other flow is detected. Calculate the total width, and if the total idle bandwidth in any path of the other flow is equal to or greater than the bandwidth request of the specific flow, reschedule the other flow and The flow scheduling method according to appendix 11, wherein one or more paths are allocated for the specific flow from the paths.

(Appendix 14)
Said assigning step comprises:
When the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request of the specific flow, the other flows are rearranged in the descending order of the total idle bandwidth in the arbitrary path and rearranged. Select one flow from the selected flows in order, and reschedule the selected flow until the total idle bandwidth in any path of the specific flow is equal to or greater than the bandwidth request of the specific flow The flow scheduling method according to Supplementary Note 13.

(Appendix 15)
Said assigning step comprises:
If the total idle bandwidth in any path of the other flow is less than the bandwidth request for the specific flow, reschedule the other flow to the idle path and set the idle bandwidth in the arbitrary path of the specific flow to: 14. The flow scheduling method according to appendix 13, wherein the flow is assigned to the specific flow and a remaining flow that still passes through an arbitrary path of the specific flow even after rescheduling based on the Maximin fairness principle.

(Appendix 16)
The flow scheduling method according to claim 10, further comprising the step of transmitting the flowlet information received from the transmission source host to a destination host.

(Appendix 17)
The flow scheduling method according to appendix 16, wherein the flowlet information includes a port number of a flowlet.

(Appendix 18)
Bandwidth reservation means configured to apply for bandwidth for a specific flow to a central management device;
Correspondence for dividing the specific flow into one or more flowlets based on the scheduling result returned from the central management unit and transferring one or more flowlets on the path allocated by the central management unit And a flow control means configured to establish a connection to be made.

(Appendix 19)
The host according to appendix 18, wherein the flow control means is configured to adjust a corresponding connection when a command to change a flowlet is received from the central management apparatus.

(Appendix 20)
The host according to appendix 18 or appendix 19, further comprising data reconstruction means configured to reconstruct data received via one or more flowlets.

4110: Forwarding table 4120: Packet forwarding means 4130: Flow statistics reporting means 4210: Switch controller 4220: Flow scheduling means 4230: Flow request prediction means 4240: Flow request table 4250: Flow statistics collection means 5110: Flow control means 5120: Bandwidth Reservation means 5210: Switch controller 5220: Flow scheduling means 5230: Bandwidth usage status / path information table 5240: Bandwidth request collection means 5310: Transfer table 5320: Packet transfer means 8110: Flow control means 8130: Data reconstruction means 8120: Bandwidth reservation means 8210: Switch controller 8220: Flow scheduling means 8230: Bandwidth usage status / path information table storage means 82 40: Bandwidth request collection means 8310: Forwarding table 8320: Packet forwarding means 830: Switch 80: Dynamic flowlet scheduling system

Claims (10)

A bandwidth request collection means configured to collect bandwidth requests for a specific flow from a source host;
Storage means configured to store link bandwidth usage and an arbitrary path of the host pair;
Based on the bandwidth usage status of the current link, one or more paths from the arbitrary path of the specific flow are allocated for the specific flow, the allocated path is returned to the transmission source host, and the specific path is transmitted from the transmission source host. Flow scheduling means configured to receive flowlet information of each flowlet obtained by dividing the flow into a predetermined number ;
Based on the scheduling result of the flow scheduling means, the central management device in which Ru and a switch controller configured to set the corresponding switch,
Bandwidth reservation means configured to apply for bandwidth for a specific flow to the central management device;
Correspondence for dividing the specific flow into one or more flowlets based on the scheduling result returned from the central management unit and transferring one or more flowlets on the path allocated by the central management unit A host comprising flow control means configured to establish a connection to
The flow control means includes
Dividing the flow into flowlets according to the number of paths allocated by the flow scheduling means, establishing a connection so that one path is allocated to one flowlet;
The flowlet information includes a port number used by the flowlet.
A dynamic flowlet scheduling system characterized by that . The flow scheduling means calculates a total idle bandwidth in an arbitrary path of the specific flow, and when the total idle bandwidth in an arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, The dynamic flowlet scheduling system according to claim 1, wherein one or more paths from an arbitrary path of the specific flow are configured to be allocated for the specific flow. The flow scheduling means rearranges the arbitrary paths of the specific flow in descending order of the idle bandwidth when the total idle bandwidth in the arbitrary path of the specific flow is equal to or greater than the bandwidth request of the specific flow, One path is sequentially selected from the rearranged arbitrary paths until the total idle bandwidth of the selected path is equal to or greater than the bandwidth request of the specific flow. The dynamic flowlet scheduling system according to claim 2. The flow scheduling means detects another flow passing through the arbitrary path of the specific flow when the total idle bandwidth in the arbitrary path of the specific flow is lower than the bandwidth request of the specific flow, and detects the other flow If the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request for the specific flow, the other flow is rescheduled. The dynamic flowlet scheduling system according to claim 2, wherein one or more paths from an arbitrary path of the specific flow are allocated for the specific flow. If the total idle bandwidth in the arbitrary path of the other flow is equal to or greater than the bandwidth request of the specific flow, the flow scheduling means determines that the other flow has a large total idle bandwidth in the arbitrary path. Reorder in order, select one flow from the reordered flows in order, and the selected flow until the total idle bandwidth in any path of the specific flow is equal to or greater than the bandwidth request of the specific flow The dynamic flowlet scheduling system of claim 4, wherein the dynamic flowlet scheduling system is configured to reschedule . The flow scheduling means reschedules the other flow to the idle path when the total idle bandwidth in the arbitrary path of the other flow is lower than the bandwidth request of the specific flow, and determines the arbitrary path of the specific flow. 5. The idle bandwidth in claim 4 is configured to allocate the idle bandwidth to the specific flow and a residual flow that still passes through an arbitrary path of the specific flow after rescheduling based on a Maximin fairness principle. The described dynamic flowlet scheduling system . The dynamic flowlet scheduling system according to claim 1, wherein the flow scheduling unit is configured to transmit the flowlet information received from the transmission source host to a destination host. 8. The dynamic flowlet scheduling system according to claim 7, wherein the storage means is further configured to store flowlet information. A flow scheduling method by a dynamic flowlet scheduling system comprising a plurality of hosts and a central management device,
Bandwidth request collecting means provided in the central management device collects a bandwidth request of a specific flow from a transmission source host,
A storage unit included in the central management device stores a bandwidth usage status of the link and an arbitrary path of the host pair;
The flow scheduling means included in the central management device allocates one or more paths from the arbitrary path of the specific flow for the specific flow based on the bandwidth usage status of the current link, and transmits the allocated path to the source A flow scheduling step of returning to the host and receiving flowlet information of each flowlet obtained by dividing the specific flow into a predetermined number from the transmission source host;
A switch controller provided in the central management device, a switch control step of setting a corresponding switch based on a scheduling result of the flow scheduling means;
Bandwidth reservation means provided in the host, a bandwidth reservation step of applying a bandwidth for a specific flow to the central management device;
The flow control means provided in the host divides the specific flow into one or more flowlets based on the scheduling result returned from the central management apparatus, and one or more on the path assigned by the central management apparatus A flow control step for establishing a corresponding connection for forwarding a flowlet of
In the flow control step,
Divides the flow into flowlets according to the number of paths allocated by the flow scheduling step, and establishes connections so that one path is allocated to one flowlet;
The flowlet information includes a port number used by the flowlet.
A flow scheduling method characterized by the above . A flow scheduling program in a dynamic flowlet scheduling system comprising a plurality of hosts and a central management device,
The bandwidth request collection means provided in the central management device executes a bandwidth request collection process for collecting a bandwidth request of a specific flow from the transmission source host,
The storage means provided in the central management device executes a storage process for storing the bandwidth usage status of the link and the arbitrary path of the host pair,
Based on the current bandwidth usage status of the link, one or more paths from the arbitrary path of the specific flow are allocated for the specific flow to the flow scheduling means included in the central management apparatus, and the allocated path is the source Return to the host and execute flow scheduling processing for receiving the flowlet information of each flowlet obtained by dividing the specific flow into a predetermined number from the transmission source host,
Based on the scheduling result of the flow scheduling means, the switch controller included in the central management device is caused to execute a switch control process for setting a corresponding switch,
The bandwidth reservation means provided in the host causes the central management device to execute a bandwidth reservation process for applying for a bandwidth for a specific flow,
Based on the scheduling result returned from the central management device to the flow control means provided in the host, the specific flow is divided into one or more flowlets, and one or more on the path assigned by the central management device To execute the flow control process to establish the corresponding connection to transfer the flowlet of
In the flow control process,
Divides the flow into flowlets according to the number of paths allocated by the flow scheduling process, and establishes connections so that one path is allocated to one flowlet;
The flowlet information includes a port number used by the flowlet.
A flow scheduling program characterized by the above .

Images