JP2018137570A - Load control system and load control method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem in which when packets are distributed to a single logical server that is virtualized with SSI techniques and communicates through a physical link group bundled with link aggregation, it is not possible to perform control on which of a plurality of physical servers the packets are distributed to.SOLUTION: A load control system S includes: an L2 switch 1 for applying link aggregation setting to a first physical link group; an open flow switch 2 for distributing packets between the first physical link group and a second physical link group; and a plurality of physical servers 3A and 3B that are each connected with any of the second physical link group and virtualized into a single virtual server 4 with SSI (Single System Image) techniques.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a load control system and a load control method suitable for load distribution of physical servers in a large-scale data center or the like.

  SSI (Single System Image) technology means that a plurality of physical servers are virtualized as one logical server and a large-scale virtual server is constructed by coordinating a plurality of server resources (Non-patent Document 1). ). As a result, the performance of the entire system is improved, high-load processing is realized, and a large amount of data can be used, which is suitable for, for example, a load balancer and Web API processing. Software that implements the SSI technology operates in cooperation with, for example, an OS (Operating System).

  Furthermore, paying attention to NIC (Network Interface Card), assigning NIC to logical servers across multiple physical servers and setting LAG (Link Aggregation) in L2 (Layer 2) switch is advantageous from the viewpoint of load distribution. large.

Bruce J. Walker, “Open Single System Image (openSSI) Linux Cluster Project”, [online], Hewlett-Packard, [Search February 10, 2017], Internet <URL: http://www.openssi.org /ssi-intro.pdf>

  The L2 switch for which LAG is set distributes the packet to a plurality of NICs. The NIC to which the packet transmitted from the client to the virtualization server is distributed depends on an algorithm set in advance in the L2 switch. The NIC to which the L2 switch assigns the packet is black boxed. In general, it is impossible to control which NIC the packet transmitted from the client to the virtualization server is distributed by controlling this L2 switch.

For this reason, it is not possible to control to which physical server a NIC is provided that distributes packets that have reached this L2 switch. Therefore, in such a system, it is not possible to distribute the load imbalance of the physical server. Note that such an uneven load on the physical server occurs, for example, when access concentrates on a specific physical server.
Further, it is not possible to stop processing on a specific physical server when updating the OS. This is because the L2 switch may distribute packets to this specific physical server.

  Therefore, an object of the present invention is to provide a load control system and a load control method that can control which physical server to which each packet is distributed.

  In order to solve the above-described problem, in the invention according to claim 1, an L2 switch that performs link aggregation setting for the first physical link group, the first physical link group, and the second physical link group, OpenFlow switch that sorts packets between the two, and a plurality of physical servers each connected to one of the second physical link groups and virtualized to a single virtual server by SSI (Single System Image) technology And a load control system comprising:

  In this way, it is possible to control to which physical server each packet received from the outside is distributed.

  In the invention according to claim 2, the physical server includes a load detection unit that detects its own load, and the virtualization server detects a load bias by the load detection unit included in each physical server, 2. The distribution change unit according to claim 1, further comprising: a distribution change unit that changes a distribution destination of packets between the first physical link group and the second physical link group with respect to the OpenFlow switch. Load control system.

  By doing in this way, it is possible to adjust the load of each physical server constituting this virtual server.

  According to a third aspect of the present invention, when the OpenFlow switch receives an instruction to change the distribution destination of a packet between the first physical link group and the second physical link group, the load of each physical server 3. The load control system according to claim 2, wherein packets between the first physical link group and the second physical link group are distributed so as to level the data.

  By doing in this way, the load of each physical server which comprises a virtualization server can be equalized.

  In the invention according to claim 4, the virtualization server does not distribute packets to a physical link connected to a predetermined physical server in the second physical link group with respect to the OpenFlow switch. The load control system according to claim 1, further comprising a process stop determination unit to be changed.

  By doing in this way, it becomes possible to stop any of the physical servers constituting the virtualization server without stopping the packet processing by the virtualization server.

  In the invention according to claim 5, the processing stop determination unit does not distribute packets to a physical link connected to a predetermined physical server in the second physical link group with respect to the OpenFlow switch. The load control system according to claim 4, wherein the predetermined physical server is stopped after a predetermined time has elapsed.

  By doing so, packet processing by the virtualization server is not erroneously stopped.

  According to a sixth aspect of the present invention, the OpenFlow switch includes a load detection unit that detects a load of each physical server, and the virtualization server detects the load bias by the load detection unit, and the OpenFlow switch The load control according to claim 1, further comprising: a distribution changing unit that changes a distribution destination of packets between the first physical link group and the second physical link group. The system.

  By doing in this way, the load of these physical servers can be detected, without providing a load detection part in each physical server.

  In the invention according to claim 7, the link aggregation setting is performed for the logical link group, and the L2 switch that distributes packets between the logical link group and the physical link group, and each is connected to the physical link group. And a plurality of physical servers virtualized to a single virtualization server by SSI technology.

  By doing in this way, the high load process by a virtual server is realizable.

  In the invention according to claim 8, the L2 switch for performing the link aggregation setting for the first physical link group, each connected to one of the second physical link groups, and a single virtual link by the SSI technology The OpenFlow switch is connected to a plurality of physical servers virtualized by the virtualization server, and the OpenFlow switch performs the load control so that the load on each physical server is leveled. The load control method is characterized by executing a step of distributing packets between the first physical link group and the second physical link group.

  In this way, it is possible to control to which physical server each packet received from the outside is distributed.

  According to the present invention, it is possible to control to which physical server each packet is distributed.

It is a lineblock diagram of the load control system in a 1st embodiment. It is a sequence diagram explaining a distribution destination change process. It is a sequence diagram explaining a stop process. It is a figure explaining the distribution operation of a comparative example. It is a figure explaining distribution operation of a 1st embodiment. It is a block diagram of the load control system in 2nd Embodiment. It is a block diagram of the load control system in 3rd Embodiment.

Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings.
FIG. 1 is a configuration diagram of a load control system S in the first embodiment.
The load control system S of the first embodiment includes an L2 switch 1, physical lines 10a to 10d, an OpenFlow switch 2, physical lines 20a to 20d, and physical servers 3A and 3B. These physical servers 3A and 3B operate as one virtual server 4 virtualized by the SSI technology. In the first embodiment, a physical OpenFlow switch 2 is provided after the L2 switch 1 to switch the packet relay destination.

  The L2 switch 1 includes a LAG setting unit 11, an interface 12, and interfaces 13a to 13d. The interface 12 is connected to an edge (not shown) of the Internet 9. The interfaces 13a to 13d are connected to the interfaces 22a to 22d of the OpenFlow switch 2 via physical lines 10a to 10d (first physical link group), respectively.

  The L2 switch 1 determines a relay destination based on a MAC (Media Access Control) address included as destination information in the packet, and relays the packet. The L2 switch 1 further bundles a plurality of physical lines 10a to 10d (first physical link group) connected to its own interfaces 13a to 13d by the LAG setting unit 11. In this way, bundling physical link groups is called ring aggregation setting. The L2 switch 1 further relays the packet destined for the virtualization server 4 to any one of the plurality of physical lines 10a to 10d bundled by the LAG setting.

  The OpenFlow switch 2 includes a distribution setting unit 21, interfaces 22a to 22d, interfaces 23a to 23d, and a flow table 25. The flow table 25 is set by the OpenFlow controller. The distribution setting unit 21 of the OpenFlow switch 2 switches the packet relay destination based on the flow table 25. The interfaces 23a and 23b are connected to the NICs 31a and 31b of the physical server 3A via the physical lines 20a and 20b (second physical link group), respectively. The interfaces 23c and 23d are connected to the NICs 31c and 31d of the physical server 3B via the physical lines 20c and 20d, respectively. Hereinafter, when the NICs 31a to 31d are not particularly distinguished, they are simply referred to as NIC31.

  The physical server 3A includes NICs 31a and 31b and a load detection unit 32. The physical server 3B includes NICs 31c and 31d and a load detection unit 32. The load detection unit 32 is for detecting a load deviation of the physical servers 3A and 3B. The load detection unit 32 detects the loads of the physical servers 3A and 3B from CPU (Central Processing Unit) resources, memory resources, NIC communication resources, and the like (not shown) included in the physical server 3, respectively. These physical servers 3A and 3B embody the virtual server 4 virtualized by the SSI technology. Hereinafter, when the physical servers 3A and 3B are not distinguished, they are simply referred to as the physical server 3.

The virtualization server 4 includes a distribution change unit 41 and a process stop determination unit 42 that function as an OpenFlow controller, and a packet processing unit 43 that processes packets. The virtualization server 4 is realized by a plurality of physical servers 3A and 3B.
The distribution changing unit 41 and the process stop determining unit 42 set the flow table 25 of the OpenFlow switch 2 when the load detecting unit 32 detects a load deviation of the physical server 3. As a result, the OpenFlow switch 2 switches the packet relay destination, and the packet distribution destination to the virtualization server 4 is changed.
The operation of the distribution changing unit 41 will be described with reference to FIG. The operation of the process stop determination unit 42 will be described with reference to FIG.

FIG. 2 is a sequence diagram for explaining the distribution destination changing process.
When the L2 switch 1 receives a packet from an external device connected to the Internet 9 (step S10), the distribution destination changing process starts. This packet is destined for the virtualization server 4. The L2 switch 1 bundles physical lines 10a to 10d for communicating with the virtualization server 4 with LAG settings. Therefore, the L2 switch 1 determines the transfer destination of this packet according to this LAG setting (step S11), and transfers this packet via the determined transfer destination path (step S12). This packet is transferred to the OpenFlow switch 2 via any of the physical lines 10a to 10d.

  When receiving this packet, the OpenFlow switch 2 determines the transfer destination of this packet according to the distribution setting (step S13), and transfers this packet via the determined transfer destination path (step S14). This packet is transferred to one of the plurality of physical servers 3A and 3B via any one of the physical lines 20a to 20d. In FIG. 2, the physical server 3 is described.

When receiving this packet, the physical server 3 transfers the packet to the virtualization server 4 by the SSI technique (step S15). The virtualization server 4 determines the physical server 3 that executes the process related to this packet (step S16), and instructs the determined physical server 3 to execute the process related to this packet (step S17). The processes in steps S15 to S17 are all performed by the SSI technique. The physical server 3 instructed to perform the process related to this packet executes the process related to this packet (step S18).
Note that the physical server 3 that transfers the packet in step S15 may be different from the physical server 3 that executes the process related to the packet in step S18. In this case, a signal including a processing instruction is transmitted from one physical server 3 to the other physical server 3.

In step S <b> 19, processing bias occurs in the specific physical server 3. The virtualization server 4 inquires each physical server 3 about the load (step S20) and receives this load response (step S21). The processes in steps S20 and S21 are processes (polling) that the virtualization server 4 periodically performs.
When the virtualization server 4 detects a load imbalance in the specific physical server 3 (step S22), the virtualization server 4 determines to change the distribution (step S23). The virtualization server 4 calculates a distribution setting that equalizes the load on the physical server 3. The virtualization server 4 transmits a request for changing the calculated distribution setting to the OpenFlow switch 2 (step S24).

  When receiving the distribution setting change request, the OpenFlow switch 2 changes its distribution setting (step S25), and responds to the virtualization server 4 with a change completion notification (step S26). Thereby, the load in each physical server 3 is leveled.

FIG. 3 is a sequence diagram illustrating the stop process.
The processes in steps S30 to S38 in FIG. 3 are the same as the processes in steps S10 to S18 shown in FIG.
For example, based on the operation of the system administrator or the like, the virtualization server 4 determines to stop the processing of the specific physical server 3 (step S39). The virtualization server 4 calculates the distribution setting so that the packet is not distributed to the physical server 3. The virtualization server 4 transmits a request for changing the calculated distribution setting to the OpenFlow switch 2 (step S40).

  Upon receiving the distribution setting change request, the OpenFlow switch 2 changes its own distribution setting (step S41), and responds to the virtualization server 4 with a distribution setting change completion notification (step S42). Thereafter, the packet cannot be distributed to the specific physical server 3.

The virtualization server 4 waits for a predetermined time from the time of step S42 (step S43). This is because even after the packet distribution to the specific physical server 3 is stopped, the physical server 3 may be processing the packet received before the stop. After step S43, the system administrator can stop the processing of the specific physical server 3 without interrupting the packet processing by the virtualization server 4.
The virtualization server 4 further cancels the SSI cluster setting of the specific physical server 3 (step S44). Thereafter, the system administrator may stop the processing of this specific physical server 3.
The virtualization server 4 may monitor the processing state of the specific physical server 3 from the time of step S42, and release the SSI cluster setting of the specific physical server 3 after the packet processing is stopped.

  Note that the OS and application of the physical server 3 may be updated after the packets are not distributed to the specific physical server 3. The system administrator can update the OS and application of a specific physical server 3 without interrupting packet processing by the virtualization server 4.

FIG. 4 is a diagram for explaining the sorting operation of the comparative example.
The load control system S of the comparative example includes an L2 switch 1, physical lines 10a to 10d, and physical servers 3A and 3B. These physical servers 3A and 3B operate as one virtual server 4 virtualized by the SSI technology.
The L2 switch 1 includes a LAG setting unit 11, an interface 12, and interfaces 13a to 13d. The interface 12 is connected to an edge (not shown) of the Internet 9. The interfaces 13a and 13b are connected to the NICs 31a and 31b of the physical server 3A via the physical lines 10a and 10b, respectively. The interfaces 13c and 13d are connected to the NICs 31c and 31d of the physical server 3A via the physical lines 10c and 10d, respectively.

  The L2 switch 1 determines a relay destination based on a MAC address included as destination information in the packet, and relays the packet. The L2 switch 1 further bundles a plurality of physical lines 10a to 10d (first physical link group) connected to its own interfaces 13a to 13d by the LAG setting unit 11. The L2 switch 1 further relays the packet destined for the virtualization server 4 to any one of the plurality of physical lines 10a to 10d bundled by the LAG setting. The broken line in the figure indicates a packet destined for the virtualization server 4. The example shown in FIG. 4 indicates that the amount of packets relayed via the physical lines 10a and 10b is large and the amount of packets relayed via the physical lines 10c and 10d is small. At this time, the load on the physical server 3A is large and the load on the physical server 3B is small.

FIG. 5 is a diagram for explaining the sorting operation of the first embodiment.
The example shown in FIG. 5 indicates that the amount of packets relayed via the physical lines 10a and 10b is large and the amount of packets relayed via the physical lines 10c and 10d is small.
Further, the distribution setting unit 21 (see FIG. 1) of the OpenFlow switch 2 switches the packet relay destination based on the flow table 25 (see FIG. 1).

  The OpenFlow switch 2 relays the packet relayed via the physical line 10a to the NIC 31a of the physical server 3A via the physical line 20a. The OpenFlow switch 2 relays the packet relayed via the physical line 10b to the NIC 31c of the physical server 3B via the physical line 20c. The OpenFlow switch 2 relays the packet relayed via the physical line 10c to the NIC 31b of the physical server 3A via the physical line 20b. The OpenFlow switch 2 relays the packet relayed via the physical line 10d to the NIC 31d of the physical server 3B via the physical line 20d. That is, the OpenFlow switch 2 relays a packet relayed through the interface 13b to the NIC 31c of the physical server 3B. The OpenFlow switch 2 further relays the packet relayed through the interface 13c to the NIC 31b of the physical server 3A. In this way, by changing the distribution destination, the amount of packets processed by the physical server 3A and the amount of packets processed by the physical server 3B are leveled.

  In the first embodiment, each physical server 3A, 3B has a plurality of NICs 31 respectively. As a result, the OpenFlow switch 2 can level the amount of packets to some extent by simply switching the connection between the one interface 22a to 22d and the other interface 23a to 23d.

FIG. 6 is a configuration diagram of the load control system S in the second embodiment.
Unlike the first embodiment, the load control system S of the second embodiment is provided with a distribution setting unit 52 that is a logical OpenFlow switch after the L2 switch 5. The distribution setting unit 52 can switch the packet relay destination.
The L2 switch 5 includes an interface 12, a LAG setting unit 51, a distribution setting unit 52, and interfaces 57a to 57d. The interface 12 is connected to the edge (not shown) of the Internet 9. The interfaces 57a to 57d are connected to the NICs 31a to 31d, respectively.
The LAG setting unit 51 includes a logical interface 53 and logical interfaces 54a to 54d. The logical interface 53 is logically connected to the interface 12. The logical interfaces 54a to 54d are logically connected to the logical interfaces 55a to 55d, respectively.
The distribution setting unit 52 includes logical interfaces 55a to 55d, logical interfaces 56a to 56d, and a flow table 58. The logical interfaces 56a to 56d are logically connected to the interfaces 57a to 57d.

  As described above, in the second embodiment, the L2 switch 5 is provided with the distribution setting unit 52 that is a logical OpenFlow switch. This also makes it possible to control which physical server each packet is distributed to.

FIG. 7 is a configuration diagram of the load control system S in the third embodiment.
Unlike the first embodiment, the load control system S of the third embodiment is provided with a load detection unit 24 in the OpenFlow switch 2. The load detection unit 24 detects the amount of each packet relayed via each interface 23a to 23d.
The load detection unit 24 may detect the amount of packets distributed to each physical server 3 and thus detect the load.

(Effect of this embodiment)
In the conventional L2 switch, the packet distribution destination cannot be changed. However, in the first to third embodiments, the packet distribution destination can be changed by adding a function to the physical OpenFlow switch 2 or the L2 switch 1. Thereby, the load control system S can eliminate the load imbalance of the physical server 3.
Furthermore, by concentrating the processing on a specific physical server or canceling the processing on a specific physical server, it is possible to update the OS and applications without interrupting the service.

(Modification)
The present invention is not limited to the above-described embodiment, and can be modified without departing from the spirit of the present invention. For example, there are the following (a) to (e).
(A) In the first and second embodiments, the physical server 3 is provided with a load detection unit 32. In the third embodiment, a load detection unit 24 is provided in the OpenFlow switch 2 to detect the load by detecting the number of packets distributed to the physical server 3. However, the present invention is not limited to this, and a logical OpenFlow switch and a load detection unit may be provided in the L2 switch. Thereby, the number of packets distributed to each physical server can be detected to detect the load on each physical server. Therefore, the number of packets to each physical server can be leveled, and the load on each physical server can be leveled.
(B) The number of each physical server is not limited to two, and may be three or more.
(C) The number of NICs that each physical server has is not limited to two.
(D) The operation of the OpenFlow switch is not limited to the operation of distributing packets between one physical link group and the other physical link group on a one-to-one basis, and may be distributed on an arbitrary basis.
(E) The virtualization server is not limited to monitoring the load of each physical server. Each physical server may monitor its own load and notify the virtualization server when the load exceeds a predetermined threshold.

S Load control system 1 L2 switches 10a to 10d Physical line (first physical link group)
11 LAG setting unit 12, 13a to 13d Interface 2 OpenFlow switches 20a to 20d Physical line (second physical link group)
21 Distribution setting units 22a to 22d, 23a to 23d Interface 24 Load detection unit 25 Flow table 3 Physical server 3A Physical server 3B Physical servers 31a to 31d NIC
32 Load detection unit 4 Virtualization server 41 Distribution change unit 42 Processing stop determination unit 43 Packet processing unit 5 L2 switch 51 LAG setting unit 52 Distribution setting unit 53, 54a to 54d, 55a to 55d, 56a to 56d Logical interface 57a ~ 57d Interface 58 Flow table 9 Internet

Claims (8)

An L2 switch that performs link aggregation setting for the first physical link group;
An OpenFlow switch that distributes packets between the first physical link group and the second physical link group;
A plurality of physical servers each connected to one of the second physical link groups and virtualized to a single virtual server by SSI (Single System Image) technology;
A load control system comprising: The physical server includes a load detection unit that detects its own load,
When the virtualization server detects a load imbalance by the load detection unit included in each physical server, the virtualization server is configured to connect the OpenFlow switch between the first physical link group and the second physical link group. A distribution changing unit for changing the packet distribution destination;
The load control system according to claim 1. When the OpenFlow switch receives an instruction to change the distribution destination of a packet between the first physical link group and the second physical link group, the OpenFlow switch is configured to level the load on each physical server. Distributing packets between one physical link group and the second physical link group;
The load control system according to claim 2. The virtualization server includes a processing stop determining unit that causes the OpenFlow switch to change a physical link connected to a predetermined physical server in the second physical link group so as not to distribute packets.
The load control system according to claim 1. The processing stop determination unit causes the OpenFlow switch to change so that packets are not distributed to physical links connected to a predetermined physical server in the second physical link group, and a predetermined time has elapsed. To stop the predetermined physical server later,
The load control system according to claim 4. The OpenFlow switch includes a load detection unit that detects the load of each physical server,
When the load detection unit detects a load imbalance, the virtualization server changes a packet distribution destination between the first physical link group and the second physical link group with respect to the OpenFlow switch. A distribution change section
The load control system according to claim 1. L2 switch that performs link aggregation setting for the logical link group and distributes packets between the logical link group and the physical link group;
A plurality of physical servers each connected to the physical link group and virtualized to a single virtualized server by SSI technology;
A load control system comprising: A plurality of L2 switches that perform link aggregation setting for the first physical link group, each of which is connected to one of the second physical link groups and virtualized to a single virtualization server by SSI technology A load control method executed by an OpenFlow switch connected to a physical server,
The OpenFlow switch distributes packets between the first physical link group and the second physical link group so that the load of each physical server is leveled;
The load control method characterized by performing.

Images