KR19990072167A - Load balancing between processors on server computers - Google Patents

Abstract

The present invention relates to a server computer having a plurality of processors 20, 30, 40 that can be connected to a network and arranged to provide services to one or more client computers connected to the network. This service includes providing an information block to the client computer 50, where the processor accesses a data storage device 120 in which data for the information block is stored. According to a preferred embodiment of the present invention, each processor in the server, in response to a request from the client computer 50 for a particular block of information from that processor, retrieves data from the data storage device 120 and requests the requested information. It has a block retrieval means 80 used to create a file representing a block. Further, load determination means 100 is provided which periodically determines activity data for a processor and includes it in the load distribution record maintained for all processors in the server. Then, for each reference to a subsequent information block in the file created by the block retrieval means 80, it is determined which processor will service the request from the client computer 50 for the subsequent information block, and the block retrieval is performed. Determination means 90 is used that is configured to include an address for that processor in a file created by means 80. Once the determining means has completed this process, the file is transferred to the client computer 50.

In a preferred embodiment, the service provided is a World Wide Web (WWW) service, which is used on parallel web servers that serve web pages to clients over the Internet. By using the techniques of the present invention, more effective load balancing is achieved across the processors of the web server, thereby preventing problems such as "toasting" of a particular processor.

Description

Load balancing between processors on server computers

In general, the demand for high performance server computers for use on networks, especially the Internet, is increasing day by day. The Internet is a huge network of many small networks that are interconnected through devices such as bridges and gateways. In theory, any computer connected to the Internet can exchange information with any other computer connected to the Internet. However, because computers often contain sensitive information, many computers and networks have security schemes that variously restrict the exchange of information between computers.

Since the Internet represents a very powerful structure for disseminating information through this, many services have been developed for using the Internet to transfer information from one computer to another. Examples of these services include FTP, News, Gopher, and the World Wide Web (hereinafter referred to as WWW).

WWW is more known and one of the fastest spreading of these services. WWW, supported by the Internet, can combine techniques such as voice, video, and text, and present them on demand, on-demand, and interactive fashion to any user connected to the Internet. Indeed, according to the WWW, anyone connected to the Internet with a computer having the appropriate software and hardware configuration can search for any HyperMedia document that can be used anywhere on the Internet. The hypermedia document is received as a file representing a block of information, within which there are hyperlinks to other documents (or blocks of information) that may include text, sound, images or movies, for example.

When a computer receiving a hypermedia document creates a document for display to a user, the computer uses some of these hyperlinks to receive other files needed to complete the display. For example, a hyperlink can be used to access an image file to be displayed as part of a document. Once the document is created by the client computer and displayed to the user, the user can request the document referenced by the hyperlink, for example by selecting a word or image on the screen using a mouse or other input device. You have the opportunity to choose.

As more people begin to use Internet services such as the World Wide Web, the demand for server computers is increasing. To meet these needs, parallel computers have begun to be used as servers, which are equipped with multiple processors to provide services. By using multiple processors, it is possible to handle very large server requests. An example of such a parallel computer is an IBM RISC System / 6000 SP machine (IBM and RISC System / 6000 are registered trademarks of IBM Corporation).

However, the problem with using such a parallel server is to balance input client traffic between processors that work together in a parallel system to serve documents (such as web pages on the WWW) and run service-based applications. . In an existing load balancing mechanism, there may be a time when a large amount of traffic is concentrated in one particular processor for a predetermined time. Expressed in technical terms, this processor is said to be "toast". In this state, the processor is taken over by the client's request and is in a virtual halt state, taking several minutes to service one request.

Accordingly, it is an object of the present invention to provide a mechanism that helps to mitigate this toasting effect by mitigating an unbalanced load on the processor.

Summary of the invention

The present invention thus provides a server computer having a plurality of processors that can be connected to a network and arranged to provide a service to one or more client computers connected to the network, the service providing information blocks to the client computer. The processor may access a data storage device that stores data for an information block, the server computer being provided to each of the plurality of processors, in response to a request from a client computer for a particular block of information from the processor, Block retrieval means used to retrieve data from the data storage device to create a file representing the requested information block, the file containing references to subsequent information blocks that may be requested by the client; and Having activity on the processor Load determination means for periodically determining the activity data and including it in a load distribution record maintained for all processors of the server, and each reference portion for subsequent blocks of information in a file created by the block retrieval means. For determining, based on the load distribution record, which processor should service a request from a client computer for its subsequent information block, and determining means configured to include an address for that processor in a file created by the block retrieval means. And delivery means for transferring the file to the client computer.

By using the approach of the present invention, the contents of the file can be dynamically changed before sending to the client, so that the contents of the file can include the address of the processor, from which address any subsequent blocks of information must be retrieved. . The determination of which processor will service a request for a subsequent block of information is made based on the load distribution record for the processor of the server. In this way, the load on the processor can be more efficiently balanced, thus mitigating problems such as "toasting" the particular processor.

In a preferred embodiment, load determination means are provided for each of the plurality of processors to determine activity data for each processor. In addition, determining means are preferably also provided for each of the plurality of processors. As an alternative to the preferred embodiment, however, one or more load determining means, determining means and transmitting means may be provided in a particular one of the plurality of processors or in a completely independent processor dedicated to performing the functions of these components. have. Thus, one processor may be dedicated to performing the function of the determining means, and all files created by the block searching means may be transferred to the dedicated processor for processing by the determining means.

In a preferred embodiment, when the client computer requests a number of subsequent blocks of information when creating the information block represented by the file, it is provided with determining means, so that the more load on the processor as indicated in the load distribution record, the more The number of subsequent information blocks to be serviced by any particular processor is determined so that the number of subsequent information blocks to service is small. Typically, one or more subsequent blocks of information will be an image file, which will be required by the client computer when creating the information block.

In a preferred embodiment, the network is the Internet, the service is a world wide web service, and the information block is a web page. However, it is apparent that the present invention can be applied to any network and service in which an information block is provided to the client that includes a reference to a subsequent block of information that the client may require and needs to be serviced by the server.

If the service is a world wide web service, the address for the subsequent information block included in the file by the determining means is preferably determined by the determining means as an identifier for the information block and a processor that will service the request for the information block. It takes the form of an address string containing all of the unique identifiers for a processor.

In a second aspect, there is provided a method of operating a server computer having a plurality of processors that can be connected to a network and installed to provide services to one or more client computers connected to the network, the service providing information to the client computer. And providing a block, wherein the processor can access a data storage device that stores data for the information block, and a method of manipulating a server computer includes: (a) a request from a client computer for a particular block of information from the processor; In response to retrieving data from the data storage device using the processor to create a file representing the requested information block, the file containing a reference to a subsequent information block that may be requested by the client. (B) all processes on the server Periodically determining activity data for the processor using load determination means for inclusion in the load distribution record maintained for the document, and (c) for each subsequent block of information in the file created in step (a). For the reference, based on the load distribution record, it is determined which processor should service the request from the client computer for its subsequent block of information and includes an address for that processor in the file created in step (a). And (d) transferring the file to the client computer.

The present invention relates to a server computer that can be connected to a network to provide services to one or more client computers, where one or more client computers are also connected to the network. In particular, the present invention relates to a server computer of a type having a plurality of processors to provide a service, and addresses the problem of balancing the load on these processors caused by a request from a client computer.

1 is a schematic block diagram illustrating a server according to a preferred embodiment of the present invention.

2 is a flow diagram illustrating a process performed in a preferred embodiment by a processor of a server when receiving a request from a client.

3 shows how load distribution records are used to determine which processor will service subsequent requests for information blocks.

In a preferred embodiment of the present invention, a mechanism for providing secondary load-balancing across multiple processors collaborating in a parallel web server, considering a WWW service provided on the Internet, is described.

Looking first at the WWW structure, it is based on Universal Resource Locators (URLs) that can be considered when specifying a specific page or file on the WWW. The URL represents the server machine and the specific file or page on that server machine. For example, the link may be "http://www.ibm.com/example.html", where "http:" indicates that you are using the World-Wide Web and "www. ibm.com "is the Internet name of a web server on the network, and the user requests a page from it, and" example.html "is the name of the page you want to receive.

The client computer uses software called a browser that provides the user with a simple GUI, allowing the user to navigate to different URLs through the WWW. Sometimes users specify a specific URL by name, but in more cases, the user will navigate from one URL to another via a reference within the page (so-called "hyperlink"). Thus, a particular word or symbol on a page may be associated with another URL, so that, for example, by clicking on the relevant location with a mouse to select the word or symbol, the client computer retrieves and displays the page from the selected URL. This new page can be located on a server that is completely different from the first page. Words or symbols that map to other URLs are sometimes referred to as "hot buttons" (or similar terms). There can be many hot buttons or hyperlinks on a single page.

Many pages or URLs can reside on a server. When the user of the client computer selects a URL on a particular server computer, the server computer performs the action specified on that page. In some cases this allows the server computer to launch the application, but in most cases the server computer retrieves the requested page from the database of stored pages and passes it back to the client computer via the Internet for display to the user.

Pages are formatted according to a standard format known as HTML. Typical web pages include textual content, embedded images (graphics), and links to a number of other pages that a user viewing the page can click to if they want more information about a particular subject. The image itself is a web "page" that a web browser can automatically request from the server computer on behalf of the client if it is downloading a page from the server computer. Typically, such images are stored as separate graphic files on the server, for example in standard ".gif" or ".jpeg" format. The HTML text includes a reference to this graphic file, which allows the client to also retrieve the specified file from the server and display both the HTML text and the associated image. The follow-on link from the page may be another page on the same server as the server that provided the base page or another web server on the Internet. To illustrate the preferred embodiment of the present invention, subsequent links on the same server as the graphics file and the base page will be mentioned. Thereafter, all of these items will be referred to as referenced Web pages.

There are a variety of commercially available web browser software for use on client computers, such as IBM's Web Explorer and NCSA's Mosaic. All of these browsers understand HTML and other WWW standard formats, and they can display or output files accurately. Similarly, mechanisms for navigating between different URLs are also understood and implemented by this program. For more information about the World Wide Web, see Andrew Ford's "Spinning the Web" (International Thomson Publishing, London 1995) and John December and Neil Randal's "The World Wide Web Unleashed." (SAMS Publishing, Indianapolis 1994).

In a preferred embodiment of the present invention, load balancing techniques are used with parallel servers so that many processors on the server appear to share the same Internet name. For example, referring to FIG. 1, a parallel web server 10, www.ibm.com, may be referred to as, for example, processors 20, 30 referred to as www1.ibm.com, www2.ibm.com, and www3.ibm.com. , 40). The load balancing software handles the request for www.ibm.com to be sent to one of the processors that form part of the parallel web server. Suitable load balancing techniques are disclosed in EP-A-0,648,038, which discloses dynamic load balancing such that the load on a parallel server is balanced across the multiple processors (or computers) forming the server. Techniques are disclosed. If a program on a client computer wants to connect to the processor of parallel server 10, it communicates with data processing system 60, often referred to as a name server, to obtain a network address for the desired server. According to the technique disclosed in EP-A-0,648,038, decision logic 70 is provided to periodically examine the processor or computer of the parallel server 10 and based on configurable criteria Select one of the processors. This configurable criterion may be selected such that the processor with the least load at the time the criterion is applied is selected by decision logic 70. The address for this processor is associated with a generic server name in storage 80 that may be used for name server 60, so that whenever the client program requests a machine address using the generic server name, decision logic 70 Is given the address of the processor in the most recently selected parallel server 10. Thus, if www.ibm.com is a generic name, the request from client 50 to name server 60 for the Internet address of www.ibm.com is such that the client is responsible for one of three processors 20, 30, 40. This will result in receiving an Internet address. For further details on this technique, reference is made to EP-A-0,648,038.

Typically, subsequent links that refer back to the same server as the main page (often referred to as 'website') will refer to the same internet name as the main page's internet name. For example, on the page http://www.ibm.com/first_page.html, the subsequent link would be http://www.ibm.com/second_page.html. Since the client computer has already asked the name server 60 for the common name www.ibm.com Internet address, it will typically use the same internet address again. Thus, subsequent requests will be directed to the same processor where the original request was serviced. No load balancing techniques used in the name server will be utilized because the name server 60 is not consulted again to resolve the generic name www.ibm.com. As a result, the actual processor that served the original request is full of requests for subsequent links, which results in the 'toasting' effect mentioned above.

The client computer can be configured to query the name server 60 for subsequent requests. However, when trying to increase the efficiency of a name server, the name server tends to store details of recent name requests in cache memory. Typically, the client will not communicate directly with name server 60. Instead, the client will generally communicate with its local name server, which is the name whose request ultimately has "authority" for name server 60 (which is parallel web server 10). Server will communicate with a series of other name servers until it is delivered. The Internet address resolved by name server 60 is then passed back through all these name servers, each of which may decide to cache a particular Internet address that is the result of the analysis of www.ibm.com. . Thus, if you configure the client to contact the name server whenever it needs to communicate with www.ibm.com, the client still tends to receive addresses that have already been resolved by the name server 60, for example For example, the local name server may determine that the client already knows the Internet address for the name and no longer needs to forward the request. Thus, it can be seen that the load balancing technique on name server 60 is useful for the client's initial request for the server address, but not for any subsequent links to the same generic server name.

To overcome this problem, it provides the parallel web server of the preferred embodiment with the ability to modify the content of the web page in the server when the web page is serviced, which uses this capability to provide the processor's ability to refer to subsequent links. You can modify the name. Considering the example of FIG. 1, since www1.ibm.com, www2.ibm.com, and www3.ibm.com are all servers within a parallel web server at www.ibm.com, otherwise www.ibm.com Subsequent links that may be referenced may be modified to refer to a particular processor, for example www1.ibm.com. This directs the request for subsequent pages to a particular processor in the parallel web server, and as apparent, the referenced web page is a graphical image embedded within the currently displayed web page or can be accessed by the user from the displayed web page. It can be applied to all referenced web pages regardless of whether they are subsequent links to different web pages.

By using the preferred embodiment's technique to specify a definite processor name (for example, www1.ibm.com) instead of the generic name (www.ibm.com) for the entire parallel server, The load can be controlled to some extent. This control can reduce the likelihood of 'toasting'.

Hereinafter, the system of the preferred embodiment will be described in more detail with reference to FIG. 1. In Figure 1, the functional components of the processor 30 required to implement the preferred embodiment are shown. Preferably all other processors 20, 40 of the server 10 that will provide WWW services include the same components, but these are omitted in the drawings for clarity. As an alternative to the preferred embodiment, one or more functional components may be provided in addition to the block retrieval means on a particular one of the plurality of processors or a completely independent processor dedicated to performing the functions of these components. Thus, one processor can be dedicated to performing the function of the determining means, and all web pages retrieved by the block searching means can be sent to the processor for processing by the determining means.

If the client computer 50 is provided with the Internet address of the processor 30 after querying the name server 60, the client 50 will directly contact the processor 30 as a request for a web page. The block retrieval means 80 will access the data in the storage device 120 to create a file representing the web page requested by the client 50. For any web page referenced in the file, determining means 90 will be used to determine which particular processor should be selected to process subsequent requests for that web page, and an identifier for that processor (e.g., www1.ibm.com) will be included at the appropriate point in the file.

As will be discussed in more detail later, the determining means 90 refers to a load distribution record that provides information about the load for all processors in the server when determining which processor to process subsequent requests. The load determination means 100 is configured to periodically evaluate the activity of the processor 30 and to include the information in the load distribution record.

Once the determining means determines which processor will process any referenced web page and the information about it is contained in a file created by the block retrieval means, the file is sent to the browser 130 of the client computer 50. Delivered to delivery means 110 for transmission. The browser 130 will then create a web page from that file for presentation to the user of the client computer 50 and send a request for the referenced web page directly to the processor identified in this file.

We now describe how to obtain a load distribution record. The parallel web server collects statistics on how busy the server's processor is, and can then use this statistics to generate a "load distribution record". Obviously, you can implement this in many ways. For example, each processor may include routines that periodically determine how busy the processor is and send that information to a predetermined central location for inclusion in the load distribution record. Load balancing records are kept on the data storage device 120 for each processor to access, or instead of shared memory accessible to all processors on the server if the server's architecture supports a shared memory structure. Can be stored in some. Alternatively, a series of messages may be sent to each processor over a communication network to periodically evaluate how busy each processor is and then send that information to all other processors in the parallel web server. Each processor can then keep a load distribution record in its local memory and update it each time it receives new load data from another processor.

The determining means 90 of each processor 20, 30, 40 then uses this load distribution record to dynamically change the Internet name of the image referenced in subsequent links and subsequently serviced web pages. This profile is updated periodically by a number configurable by the web server administrator. By making this change, it is possible to obtain the effect that the subsequent request to the processor which is getting busy becomes proportionately smaller, thereby avoiding the severe performance impact when the processor is "toasted". Once the processor is restored, more demand will be placed on the processor, thus continuing the monitoring and load-tunning cycles.

When receiving a request from a client, a process executed by a processor according to a preferred embodiment of the present invention is described below with reference to FIG. In step 200, the processor waits for a request from the client. In step 210, a request 205 is received from a browser of a client computer. Then, after the processor determines which web page is requested in step 220, the block retrieval means 80 retrieves the page from the database 120 and creates a file (step 230). Next, in step 240 any reference to another web page in the file is identified, and the determining means 90 uses the most recent load distribution record 270 to display that web page for each web page. Determine which processor should service the request. Thus, if the subsequent request on the web page is called 'second_page.html' and the determining means determines that the processor 20 should service the request, then the reference in the file is 'http: /www1.ibm.com/second_page' will be changed to read .html '.

Once this subsequent hyperlink is completed by the determining means, the file is transferred to the client browser in step 250. Next, in a preferred embodiment, the processor is configured to reevaluate how busy the processor is in step 260 and to update the load distribution record based on that estimate. After executing this, the processor returns to step 200 to wait for the next request from the client.

As mentioned above, subsequent web pages referenced in a file can be classified into two types. In the first type, you can refer to a web page that the browser needs to search while creating the current web page for display to the user. For example, a base file may contain all the text for a web page, but data for any image to be displayed will typically not include, but rather a reference to the web page containing the image is returned to the browser. Will be provided in the file. The browser will need to retrieve these web pages in order for the user to play the complete web pages.

The second type of reference web page in the file will be subsequent links that are not required by the browser to generate the initial web page for display. Instead, they will be pages that the user can select once the web page is displayed. Typically, a user will be able to select a word or image that is particularly highlighted on the screen, and these words and images are associated with other web pages. In this selection, the browser will retrieve from that file the address of the processor that will service the request for the web page and contact the processor directly.

In particular, with respect to subsequent web pages of the first type identified above, FIG. 3 shows how the determining means 90 can use the load distribution record to split the request for this first type of web page among the available processors of the server. It illustrates if there is. Because the browser will require this first type of web page almost immediately to create a complete web page for display to the user, it can be expected that these requests will be sent back to the server immediately after the original web page is delivered to the browser. have.

In Figure 3, the load on each processor is shown for each processor as a percentage of the maximum load. Typically the maximum load is determined as the point at which the processor will not be able to provide an acceptable response time, which is generally determined by the server administrator. Obviously, there are other ways to assess the load. It can be seen from FIG. 3 that processor 2 is the most loaded, processor 3 is next, and processor 1 is the least loaded. For example, if there are nine subsequent types of first web pages, the determining means may use the percentage load plot for each processor to determine how many of these nine web pages should be served by each processor.

As shown in Fig. 3, the unused load for each processor is divided into blocks of the same size, so that a total of nine blocks are generated. A suitable way to balance the load placed on a processor is to assign the responsibility to a particular processor to service subsequent requests for multiple web pages by the number of blocks associated with that processor's unused load. Thus, in the example shown in FIG. 3, the address of processor 1 is referenced for five web pages, the address of processor 2 is referenced for one web page, and the address of processor 3 is referenced for three web pages. Will be.

From the foregoing description, it is clear that according to the preferred embodiment of the present invention, in order to change the Internet name of the server processor for which any subsequent link should be searched, the contents of the web pages are changed dynamically when they are served. This processor is selected from a set of processes that can access the same web content, and this selection is made based on the relative load of these servers. In this way, more efficient load balancing is achieved and the problem of processor "toasting" is avoided before it becomes more severe.

Claims (13)

A server computer having a plurality of processors 20, 30, 40 that can be connected to a network and arranged to provide services to one or more client computers connected to the network, the service providing information blocks to the client computer 50. Wherein the processor accesses a data storage device 120 in which data for the information block is stored; A file provided to each of said plurality of processors, in response to a request from a client computer 50 for a particular block of information from the processor, retrieving data from said data storage device 120 and representing said requested block of information; Block retrieval means (80) used to create a file, the file containing references to subsequent blocks of information that may be required by the client (50); (2) load determination means 100 for periodically determining activity data for the processor and including it in a load distribution record maintained for all processors of the server; (3) For each reference portion for the subsequent information block in the file created by the block retrieval means 80, a processor requests a request from the client computer 50 for the subsequent information block based on the load distribution record. Determining means (90) configured to determine whether to service and to include an address for the processor in the file created by the block retrieval means (80); (4) a server computer comprising transfer means (110) for transferring said file to said client computer (50). The method of claim 1, The load determining means (100) is provided to each of the plurality of processors to determine activity data for each processor. The method according to claim 1 or 2, And said determining means (90) is provided to each of said plurality of processors. The method according to any one of claims 1 to 3, When creating the information block represented by the file, if a large number of subsequent information blocks are required by the client computer 50, the determining means 90 causes the processor, as indicated by the load distribution record. The server computer configured to determine the number of these subsequent information blocks to be serviced by any particular processor such that the more load on the phase, the less the number of subsequent information blocks to service by the processor. The method according to any one of claims 1 to 4, At least one of the subsequent information blocks is an image file to be requested by the client computer (50) when creating the block information. The method according to any one of claims 1 to 5, And said network is the Internet. The method of claim 6, The service is a World Wide Web service, and the information block is a web page. The method of claim 7, wherein The address for the subsequent information block included in the file by the determining means 90 is the identifier for that information block and for the processor that the determining means 90 has determined as the processor to service the request for that information block. Server computer in the form of an address string containing all unique identifiers. A method of manipulating a server computer having a plurality of processors 20, 30, 40 that can be connected to a network and arranged to provide services to one or more client computers connected to the network, the service being directed to a client computer 50. Providing an information block, wherein the processor accesses a data storage device 120 in which data for the information block is stored; (a) in response to a request from a client computer 50 for a particular block of information from a processor, in order to create a file representing the requested block of information, the data from the data storage device 120 using that processor; Retrieving a file, the file containing a reference to a subsequent block of information that may be required by the client 50; and (b) using load determination means 100 to periodically determine activity data for the processor and include it in the load distribution record maintained for all processors in the server; (c) For each reference to a subsequent block of information in the file created in step (a), a processor, based on the load distribution record, requests a request from the client computer 50 for that subsequent block of information. Determining whether to service and including an address for the processor in the file created in step (a); (d) transmitting the file to the client computer (50). The method of claim 9, Wherein each of said plurality of processors comprises load determining means (100) arranged to perform said step (b). The method according to claim 9 or 10, Each said processor comprising determining means (90) arranged to perform said step (c). The method according to any one of claims 9 to 11, If a number of subsequent blocks of information are required by the client computer 50 when creating the information block represented by the file, the step (c) indicates that the load on the processor is reduced, as indicated by the load distribution record. Determining the number of these subsequent information blocks to be serviced by any particular processor, such that the more, the less the number of subsequent information blocks to be serviced by that processor. The method according to claim 9, The address for the subsequent information block included in the file in step (c) is an identifier for that information block and a unique identifier for the processor determined as the processor to service the request for that information block in step (c). A method of operating a server computer that takes the form of an address string that includes all of the above.