JP2015231110A - Cpu allocation device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To determine core allocation for efficient table processing and to speed up packet processing in a multi-table.SOLUTION: A CPU allocation device acquires at least one of search table information sets including a use memory used by each table, an expected search time required for searching each entry in the table and an expected packet processing time for packet processing. Based on the acquired search table information, the CPU allocation device allocates to each core at least one of table search and packet processing in a manner to have the same processing time in each core, outputs an allocation result, and based on the allocation result, performs the change and the control of a content to be processed in the core.

Description

  The present invention relates to a CPU allocation apparatus, method, and program, and in particular, efficient processing for multi-table flow retrieval and packet processing in a multi-core CPU environment, as represented by OpenFlow (registered trademark) Wire Protocol. The present invention relates to a CPU allocation apparatus, method, and program for performing allocation.

  In recent years, the use of flow-based networking that enables communication quality control and path control with finer granularity by performing communication control such as OpenFlow (registered trademark) Wire Protocol in units of flows has been spreading. A flow is identified by a MAC (Media Access Control) address, a VID in a VLAN (Virtual LAN), an IP (Internet Protocol) address, a port number, or the like. A series of communication for each application, and flow-based networking enables QoS control, packet header rewriting, and path control according to the user and application.

  A flow switch, which is a relay node for flow-based networking, refers to an entry registered in the flow table and rewrites a packet header and sends a packet according to the entry. The flow entry is composed of entry priority, a matching condition having fields corresponding to each information of the receiving port and header, and processing contents to be performed on a packet that matches the matching condition.

  A flow switch corresponding to a recent protocol represented by OpenFlow (registered trademark) Wire Protocol can handle a plurality of flow tables (multi-tables) in order to improve the simplicity of the flow table. As shown in the center of FIG. 1, the flow search and packet processing are repeated in order from “table 0”.

  Until now, in the multi-table processing in the flow switch, as shown in FIG. 2, one core performs all table processing for one packet.

  As a simple parallelization, as shown in FIG. 3, a method of dividing a table to be processed for each core based on the number of tables can be considered.

  Also, the table entries are divided into those with Mask and those without Mask, those without Mask are performed by the CPU with hash, those with Mask are NIC (Network Interface Card) and GPU (Graphics Processing Unit)) There are known methods for offloading processes (see, for example, Non-Patent Documents 1 and 2).

Sangjin Han, Keon Jang, KyoungSoo Park, and Sue Moon. Packetshader: a gpu-accelerated software router. ACM SIGCOMM Computer Communication Review, 40 (4): 195-206, 2010. Voravit Tanyingyong, Markus Hidell, and Peter sjodin.Inproving pc-based Openflow switching performance.In Proceedings of the 6th ACM / IEEE Symposium on Architectures for Networking and Communications Systems, page 13.ACM, 2010.

  However, as in the example of FIG. 2, in a method in which one core performs lookup of all tables, cache misses in the CPU cache increase and the search speed deteriorates.

  In addition, parallel processing based on the number of tables as in the example of FIG. 3 increases the CPU idle time (core waiting time of FIG. 4) and cannot achieve efficient parallel processing.

  Further, the method of dividing an entry having a mask and an entry not having a mask is not different from a method in which an entry is not divided when most entries do not have a mask.

  These problems are caused by the fact that the time for searching for a flow performed in each table and the processing time for a packet are not constant, and that cache misses increase when a plurality of tables are processed in one core.

  The present invention has been made in view of the above points, and it is an object of the present invention to provide a CPU allocation device, method, and program capable of determining core allocation for efficient table processing and accelerating packet processing in a multi-table. To do.

According to one aspect, a CPU allocation device that allocates processing to each core of a flow switch having a multi-core CPU,
Obtaining at least one of used memory used by each table or information of a search table including an expected search time necessary for searching each entry of the table or an expected packet processing time for performing packet processing; A processing core management means for allocating at least one of table search and packet processing to each core based on the information of the search table so that the processing time of each core is the same, and outputting an allocation result;
There is provided a CPU allocation device having processing core control means for changing and controlling processing contents performed by the core based on the allocation result.

  According to one aspect, by distributing the table processing evenly to each core and increasing the efficiency of parallelization, flow search processing and packet processing of the flow switch can be performed at high speed, and the flow switch can be speeded up. Can be planned.

An example of conventional multi-table processing. An example in which one core performs lookup of all tables in conventional multi-table processing. The example which divides | segments a table on the basis of the number of the conventional tables. Example of table processing time and core waiting time during simple allocation to a core. The figure which shows the process outline | summary in one embodiment of this invention. The example of the core at the time of equally allocating to CPU in one embodiment of this invention. The system configuration example in one embodiment of this invention. The example of the search table in one embodiment of this invention. The figure for demonstrating the unit of the process division of the table in one embodiment of this invention. The example of the division | segmentation of the packet processing in one embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  First, the outline of the present invention will be described.

  FIG. 5 shows an outline of processing in one embodiment of the present invention.

  The present invention acquires or estimates the flow search processing and packet processing time and memory usage performed in each table of the multi-table, and performs flow search and packet processing so that the load is evenly distributed to each CPU. Is assigned to each core. By acquiring the memory usage amount, when the table search is performed, the flow search and processing can be performed at a high speed when it fits in the cache memory of the CPU. As a result, search and processing using the CPU cache effectively can be realized.

  By evenly distributing the load to each CPU, the idle time of each core is reduced as shown in FIG. 6, and the amount of packets that can be processed per unit time also increases. Since it is difficult to divide the processing time completely evenly, it is determined that the load can be distributed when the processing is allocated according to the present invention rather than the conventional simple processing allocation in units of tables. Of course, it is desirable to make them as close to each other as possible.

  For example, the processing time of each table in the examples of FIGS. 4 and 6 is as follows.

table 0 28
table 1 10
table 2 10
table 3 10
table 4 14
table 5 14
In FIG. 4 and FIG. 6, processing is performed by three cores. However, when x packets are processed, if x is sufficiently large ((x + 2) / x = 1 can be considered as large) ) The packet processing time t per packet is as follows. Here, T (core _x ) t is the packet processing time in core x, and T (table _i ) is the processing time in talbe _i .

  The processing time at the time of the simple parallelization of FIG. 4 which is the prior art is as follows.

T (core _a ) = T (table ₀ ) + T (table ₁ ) = 38
T (core _b ) = T (table ₂ ) + T (table ₃ ) = 20
T (core _c ) = T (table ₄ ) + T (table ₅ ) = 28
T = max (T (core _a ), T (core _b ), T (core _c )) × (x + 2) / x
x is big enough so
t = 38
It becomes.

  The processing time of each core at the time of the equal allocation in FIG. 6 of the present invention is as follows.

T (core _a ) = T (table ₀ ) = 28
T (core _b ) = T (table ₁ ) + T (table ₂ ) + T (table ₃ ) = 30
T (core _c ) = T (table ₄ ) + T (table ₅ ) = 28
T = max (T (core _a ), T (core _b ), T (core _c )) × (x + 2) / x
x is big enough so
t = 30
It becomes.

  In the above example, it can be seen that the processing can be performed approximately 1.27 times faster when the processing time of each core is equalized than when simple parallel processing (FIG. 4).

  Hereinafter, the processing of the present invention will be described in detail.

  FIG. 7 shows a system configuration example according to an embodiment of the present invention.

  The system shown in FIG. 1 includes a flow rule input unit 10, a search table management unit 20, a search table 30, a flow search / packet processing unit 40, and a CPU allocation device 100.

  The search table management unit 20 generates the search table 30 based on the flow rule acquired from the flow rule input unit 10. FIG. 8 shows an example of the search table 30. Each value shown in the figure is one of “value”, “value / bitmask”, and “*”, and “*” represents Don't care. Other header fields can also be used in the search, and are not limited to header information (for example, part of the payload).

  The CPU allocation device 100 includes a processing core management unit 110 and a processing core control unit 120.

  The CPU allocation device 100 is necessary for searching the memory used by each table of the created search table 30 and each entry of each table from the search table management unit 20 based on the flow rule input from the flow rule input unit 10. The information of the search table 30 such as the expected search time and the expected packet processing time for performing packet processing is acquired, and the processing allocation of the core in the flow search / packet processing unit 40 is performed using the information.

  The processing core management unit 110 acquires information on the search table 30 such as the memory used by each table and the expected search time required to search each entry of each table from the search table management unit 20, and the flow Which table and which process is to be performed in each core that performs search and packet processing is determined, and this information is transmitted to the processing core control unit 120 as an argument of a function in the program.

  The processing core control unit 120 outputs the processing contents of the flow search and packet processing performed by each core to the flow search / packet processing unit 40 according to the information acquired from the processing core management unit 110.

  The flow search / packet processing unit 40 changes processing performed in each core according to these contents.

  Hereinafter, the processing of the processing core management unit 110 and the processing core control unit 120 will be described in detail.

  The processing core management unit 110 determines which processing is to be performed in each core that performs packet processing from the contents of the acquired search table 30, and outputs the processing core control unit 120 as a flow search / packet processing and core allocation list. . The processing core determination operation will be described later.

  The processing core control unit 120 changes the content of processing performed in each core to the flow search / packet processing unit 40 according to the flow search / packet processing and core allocation list determined and input by the processing core management unit 110. ,Control.

  Hereinafter, the processing core determination operation of the processing core management unit 110 will be described.

  The processing core management unit 110 divides the processing contents of the search table 30 and determines a process to be assigned to each core based on the divided result. There are the following two methods for dividing the processing contents of the search table.

(a) Division of “flow search processing” and “packet processing”;
(b) Search table division;
Combine (a) and (b) above, and divide into entries that take time for flow search or packet processing and entries that do not, and assign different cores to entries that take time for packet processing by flow search and packet processing There is a method (FIG. 9A). Here, the time-consuming entry is an entry described as packet processing that is performed when a plurality of packet header rewrites, protocol processing, header attachment / detachment by a tunneling protocol, and the like match the entry. An entry in which packet processing as described above is described in the search table 30 can be determined as an entry that takes time. Further, as shown in FIG. 9B, there is a method of distributing the flow search process to “core a” and the packet process to “core b”.

  Also, if a flow search is performed and an entry that takes a long time to process is matched, the flow search is performed by the own core (the core that performed the search), and another core performs packet processing. In some cases, in addition to the flow search, there is an allocation method in which packet processing is performed by the own core (FIG. 10). Here, “processing takes time” means rewrite processing of n (n ≧ 1) packet headers, protocol processing, change of tunneling protocol header, complicated matching condition (high bit number to be compared, bitmask Is specified). In the example of FIG. 10, when “core a” includes “table 0” in which flow search processing and light packet processing are specified, flow search processing and light packet processing (table 0) are performed by “core a”. The heavy packet processing (talbe 0-b), which takes time, is distributed to be performed by “core c”.

  The conventional process dividing method has been considered only in units of tables, but by dividing the above two types of processes of the present invention, the unit of processing is changed from table units to flow search processes and packet processing units. It can be made smaller, and the processing can be distributed more evenly.

  Furthermore, since the frequency of processing is different for each table and for each entry, it is possible to weight each processing divided in consideration of these, and to determine processing allocation for each core based on the result.

  For example, the flow search processing and packet processing time for each table can be determined according to the following criteria.

(a) The processing time of the entry that takes the longest processing time among the entries;
(b) Average processing time for each process of each entry;
(c) The processing time of each entry multiplied by the probability that the arrival packet hits each entry;
The weighting weight is, for example, the probability that the arrival packet (c) hits the entry. The probability that an arrival packet hits an entry can be estimated in advance by some method or operated to some extent, acquired at that time, the number of hits of the entry is acquired, and the probability is calculated.

  As described above, the processing core control unit 120 controls the flow search / packet processing unit 40 to set the table for performing the processing assigned by the processing core management unit 110 to the core.

  Note that the operation of the components of the CPU allocation device 100 shown in FIG. 7 can be constructed as a program and installed in a computer used as the CPU allocation device for execution, or distributed via a network.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

DESCRIPTION OF SYMBOLS 10 Flow rule input part 20 Search table management part 30 Search table 40 Flow search and packet processing part 100 CPU allocation apparatus 110 Processing core management part 120 Processing core control part

Claims (7)

A CPU allocation device that allocates processing to each core of a flow switch having a multi-core CPU,
Obtaining at least one of used memory used by each table or information of a search table including an expected search time necessary for searching each entry of the table or an expected packet processing time for performing packet processing; A processing core management means for allocating at least one of table search and packet processing to each core based on the information of the search table so that the processing time of each core is the same, and outputting an allocation result;
Based on the allocation result, processing core control means for changing and controlling the processing content performed in the core;
A CPU allocation device comprising: The processing core management means includes:
The processing content of the search table is divided into either flow search processing or packet processing, or a combination of flow search processing and packet processing based on the information of the search table, and each core based on the divided result The CPU allocation apparatus according to claim 1, further comprising an allocation unit that determines a process to be allocated to the CPU. The assigning means includes
Means for dividing the flow search processing or packet processing into an entry that takes time and an entry that does not take time,
Means for allocating an entry that does not take a long time for the flow search process or the packet process so that the process is performed by the original core a, and the entry that takes a long time is allocated so that the process is performed by a core b different from the core a ,
The CPU allocation device according to claim 2, comprising: The assigning means includes
Means for weighting the process according to the frequency of performing the process for each table and for each entry;
Means for dividing the processing of the search table based on the weight and assigning the divided processing to each core;
The CPU allocation device according to claim 2, comprising: The weight is
The processing time of the entry that takes the longest to process,
The average processing time for each process for each entry,
5. The CPU allocation device according to claim 4, wherein one of processing times of each entry is multiplied by a probability that the arrival packet hits each entry. A CPU allocation method in an apparatus for allocating a table for processing to each core of a flow switch having a multi-core CPU,
Obtaining at least one of used memory used by each table or information of a search table including an expected search time necessary for searching each entry of the table or an expected packet processing time for performing packet processing; A processing core management step of assigning at least one of table search and packet processing to each core based on the information of the search table so that the processing time of each core is the same, and outputting an allocation result;
Based on the allocation result, a processing core control step for changing and controlling the processing content performed in the core;
CPU allocation method characterized by performing. Computer
6. A CPU allocation program for causing a CPU allocation device according to claim 1 to function as each unit.

Images