Architecture for efficient utilization and optimum performance of a network

[0001] The present invention relates to increasing performance of Wide Area Network (WAN) communications and in particular to a device positioned at both ends of a constrained link that reroutes outgoing packets to a proxy application that implements data compression.

[0002] The growth in data communication traffic, including email, client/server applications, multimedia applications, Internet and intranet applications, has continued to cause critical bandwidth shortages across many networks. This demand for instant data communication often exceeds the amount of bandwidth available, resulting in congestion and delay. As more applications move from Local Area Networks (LANs) to Wide Area Networks (WANs) application performance and response time issues become critical to continued smooth operation of an enterprise. With offices widely distributed around the globe and present day budgetary constraints, the implementation of additional wide area links is cost prohibitive for many international installations. Consequently, application and network architects need additional solutions to help them efficiently use existing bandwidth to support more applications and more end users.

[0003] The Open Systems Interconnection (OSI) reference model is a widely used model describing the flow of data traffic across a network. The OSI model has seven layers; each of the seven layers communicates with a layer below it through a specific interface and its peer layer on a different system in the network through a specific protocol. The combination of all networking protocol layers is often referred to as the networking stack. Packet based Transmission Control Protocol over Internet Protocol (TCP/IP) is perhaps the most widely known protocol in use in the Internet. IP is a network layer (Layer 3) protocol that defines a set of standards for addressing and routing of packets across a connectionless network. TCP is a connection oriented, transport layer (Layer 4) protocol that defines a way to ensure reliable transport of information.

[0004] Encoding or compression of network communication data is a natural choice for reducing the amount of traffic needed to be carried over a network link. Different schemes have been previously proposed for reducing traffic via compression. Most of these implementations concentrate on compressing traffic at the IP layer (Layer 3) of the OSI model. These Layer 3 encoding techniques are typically based on intra-packet and/or inter-packet compression schemes that have been described in the prior art. The aim of these schemes is to reduce the number of bytes transmitted in every packet; thus, these systems encode each packet or datagram independently, based upon information within each packet. Since they are not capable of detecting repetition of entire packets or repetitions that exist across multiple packets, they typically may miss many possible repetitions and therefore do not yield a high degree of reduction.

[0005] It is also well known to compress data at a much higher layer of the communications networking stack, such as at the application layer (Layer 7). For instance, images can be compressed in a variety of formats such as the Graphics Interchange Format (.gif) or the Joint Photographic Experts Group format (.jpeg). These data file-encoding formats reduce the space required for storage for such an image as well as its transmission. Hence, at the application layer a server may encode all of its images in the (.gif) format before transmission to a client on the other end of a WAN connection. Each image received by the client would then be decoded at the application layer to generate the original viewable image.

[0006] However, this form of application layer encoding and/or compression is typically performed offline, on single objects, prior to transmission. Such application layer encoding is not performed in an online fashion within a continuous communications stream. These types of encoding schemes are thus largely based only on intra-object information, that is, information contained in a single image, rather than inter-object information that consider data from all of the different objects to be transmitted from one or more servers to one or more clients.

[0007] For example, International Patent Publication Number WO 01/37516 describes a technique whereby a private network connection is implemented between devices known as accelerator exchange servers. A corresponding accelerator exchange client software is installed at a client's. The accelerator exchange client software may be implemented as a browser program adds on, for example.

[0008] U.S. Pat. No. 5,657,452 also discloses a method for setting up a connection over a data network. A proxy engine is used to implement data compression. However, this approach requires installation of the proxy engine so that it runs in a local endpoint node application.

[0009] The present invention provides an architecture for a network optimization computing device that provides bandwidth saving and performance boosting over constrained Wide Area Network (WAN) links. The network acceleration device is positioned at each one of at least two edges of a constrained WAN link. The device intercepts outgoing Internet Protocol (IP) packets, rerouting them to a proxy application on the same device. The proxy application uses a persistent connection with a proxy application located within a corresponding network accelerator device at the other end of the WAN link. The persistent connection is used to transmit the intercepted data after compressing it using a dictionary-based compression algorithm. The device uses techniques such as address spoofing and a proxy-to-proxy communication protocol to assure network transparency from the perspective of the client matrices, so that they do not have to install or run any special application software.

[0010] In another aspect, the present invention relates to the ability to intercept network traffic at a connection layer, and route it to a private, constrained WAN link. Rerouting of messages is done in a way in which it is transparent to the end nodes. That is, no specialized browser program and/or application software needs to be installed at the end user client computers. These devices only view a standard virtual connection; no modification is needed to their standard protocol layer processing. Responsibility for intercepting LAN TCP connection request packets and completing the proxy connections is entirely the function of the network accelerator devices.

[0011] This architecture provides a number of benefits for network optimization as compared to traditional packet based compression schemes.

[0012] To begin with, most other approaches concentrate on reducing the number of bytes each particular packet has. The approach of the present invention actually reduces the number of packets. This is because compression occurs at a higher layer than the TCP/IP layer, and the compressed stream is divided at inter-packets at the TCP layer. This technique thus results in a smaller number of “fat” or full packets to be transmitted over the LAN link, rather than a large number of “thin” or small packets.

[0013] In addition, data is transferred over the WAN link via a persistent connection. Thus, connection and termination requests packets are not repeatedly transmitted over the WAN link.

[0014] By reducing the overall number of packets network utilization is actually increased. This reduces the need for packet retransmissions due to errors, in turn reducing the load on routers in the WAN. As a result, a significant boost in the performance of applications running over the WAN is realized.

[0015] An architecture according to the present invention also offers higher bandwidth saving over traditional packet based compressions schemes. There are a number of reasons for this.

[0016] Since the architecture results in a significant reduction in the number of packets, sophisticated compression algorithms can be used with essentially no effect on the overall performance. Packet-based compression schemes are typically quite sensitive to the complexity of the compression algorithms employed, and tend to use relatively easy to implement and simple compression schemes. However, the significant packet reduction available with the present invention justifies using complex compression schemes, since the gain more than compensates for the extra processing time.

[0017] The invention may also use the same dictionary to compress all streams of data belonging to a given connection. Typically, different streams of the same connection will have a common context. In this instance, therefore, the dictionary will be relevant.

[0018] Finally, the invention employs a persistent dictionary for every persistent connection. Thus, all connections relevant to a particular type may have a common context. The dictionary thus becomes more or less like a cache for particular traffic types. For example, emails originating from a particular site may have the same fixed header or a trailer containing a confidentially notice. When these emails are sent from different client computers but all originating from the same end of a WAN connection, they will have many words in common. The application of a persistent common dictionary to the persistent connection between the two ends of the WAN link therefore results in efficient compression for the traffic, which is repeatedly sent between the two ends, regardless of the client computer from which it originates.

[0024] A description of preferred embodiments of the invention follows.

[0025] To achieve the desired performance improvement the present invention employs a network accelerator device (NET ACCEL) 14 at each end of a constrained Wide Area Network (WAN) connection. In the example in FIG. 1, improved communication is to be provided between a first or Local Area Network (LAN) 11-1 and a second LAN 11-2. The respective LANs 11 each consist of a number of client computing devices located at network end nodes such as devices 10-1-1, 10-1-2, . . . 10-1-m that comprise the first LAN 11-1, and similar client devices located at nodes 10-2-1, 10-2-2, . . . 10-2-n located within the second LAN 11-2. It is important to note that the devices herein referred to as the “clients” 10 are unaware that their traffic is being communicated via the network accelerators 14-1 and 14-2. Traffic passing through the network accelerators 14 is compressed in a way in which is transparent to the end nodes 10 while achieving the required bandwidth reduction. The manner of implementing this will be described in detail shortly.

[0026] In general, clients 10-1 associated with first LAN 11-1 connect one or more switches 12-1 to the network accelerator 14-1 to a router 16-1. Router 16-1 has available connections to the second LAN 11-2 through a private WAN 20 that may, for example, be Internet Protocol (IP) based. The second LAN 11-2 similarly consists of a router 16-2, network accelerator, 14-2 switches 12-2 and associated clients or nodes 10-2.

[0027] The network accelerators 14-1 and 14-2 provide a proxy server for connections established between the respective LANs 11 that they serve. Thus, as shown in FIG. 2, from the perspective of client 10-1-1 and 10-2-2, they have directly established a connection in the usual way and the existence of the proxy connection is entirely transparent to them. The implementation of such a proxy is done with known address spoofing techniques to assure transparency.

[0028] Referring now to FIG. 3 as well as FIG. 1, consider that one of the clients 101-1 in the first LAN 11-1, known as Machine A, wishes to establish a connection with another client 10-2-2 in the second LAN 11-2, known as Machine B. The interaction of the main components of the system will now be described in detail. In a first step 100, a connection request packet is transmitted from Machine A. The connection requests that a connection be established between Machine A and Machine B. The connection request may, for example, specify port x for Machine A and port y for Machine B. At the TCP level, the connection request may take the form of a SYN message.

[0029] In the next step 102, the network accelerator 14-1 associated with the first LAN 11-1 is the first to intercept the connection request. It completes the connection request with Machine A by spoofing Machine B. For example, a response by network accelerator 14-1 is provided to Machine A using the destination address and port specified in the intercepted connection address, and replying to Machine A with a proxy acknowledgement in such a way as to fool Machine A into thinking it is connecting directly to Machine B when in fact it is not. This interception is performed by a proxy application running on the network accelerator as will be described in connection with FIG. 4.

[0030] The proxy application running on network accelerator 14-1 then assigns one of the persistent connections it has with the network accelerator 14-2 to handle the connection requested by Machine A. This can be done through the process beginning at state 104.

[0031] For example, a determination can first be made by network accelerator 14-1 if port y is reachable through an available persistent connection. If so, the existing connection with the second network accelerator 14-2 associated with the second LAN 11-2 will be discovered, and the connection request by Machine A will be associated with this existing persistent connection. If, however, no suitable persistent connection already exists, then a state 106 is entered in which a new persistent connection will be requested and established between network accelerators 14-1 and 14-2. This can be accomplished by passing connection request messages through the network that discover other compatible network accelerator 14.

[0032] In the next state 108, once the new persistent connection is established a determination is made whether or not a persistent connection has already been established for the same traffic type. If the answer is yes, then a copy of the existing compression dictionary for that persistent connection will be assigned to handle the new connection in state 110.

[0033] In any event, processing next continues to a state 112 where the new connection information is passed in a message between network accelerator 14-1 and network accelerator 14-2. The information defines the characteristics of the desired connection between Machine A and Machine B.

[0034] In state 114, network accelerator 14-2 finally receives a connection request. In response, it then sends its own connection request to Machine B on its local LAN 14-2. This connection request is established at port y using a source address for Machine A and source port x. Thus, network accelerator 2 also spoofs its connection to Machine B at its local end. With the end to end connection now set up through the proxies associated with network accelerators 14-1 and 14-2, LAN packets may now travel between Machine A and Machine B through the proxies provided by network accelerators 14-1 and 14-2. All packets related to established connections are intercepted by a network accelerator 14 and rerouted to a proxy application running on it. The proxy application compresses the rerouted data using a dictionary assigned to the persistent connection that the data belongs to. The compressed data is then sent to the remote network accelerator at the other end of the proxy connection. The proxy running on the remote network accelerator decompresses the received stream, and then sends it to the corresponding client using the source and destination address and ports that it has for this connection.

[0035] If the clients 10 are using Datagram Protocol (UDP), the traffic is handled in a slightly different way. Generally, it is desirable if UDP packets are intercepted and compressed as explained above. However, they are preferably retransmitted between the network accelerators via a suitable persistent TCP connection. Moreover, it is possible to parse some UDP-based protocols, such as RTSP, in order to obtain information that can facilitate compression.

[0036] FIG. 4 is a high-level software diagram for implementation of the invention. An IP packet routing module within each network accelerator 14 performs packet redirection functions on incoming LAN packets. These are passed through IP and TCP layers, redirecting the packets to a proxy application 200. The proxy application 200 may access the rerouted data via standard socket API calls.

[0037] The proxy application then receives, compresses and redirects data to an appropriate proxy connection (as was described in connection with the steps 104 through 114 in FIG. 3.) On the receiver side, compressed data is fed out from the proxy application, back down through the protocol layers to provide the outgoing proxy packets.

[0038] The system therefore consists of at least two network accelerators 14-1 and 14-2 with one positioned at each end of a Wide Area Network (WAN) link. The WAN link provides available persistent connections between network accelerator machines 14.

[0039] In order for each remote network accelerator to be informed of the characteristics of the connection it is dealing, a proxy-to-proxy protocol is employed. Information transmitted via this proxy-to-proxy protocol includes at least the original transport protocol i.e., information as to whether or not the original protocol is TCP or UDP, original addresses and parts, start and end points for data and any possible error conditions.

[0040] In addition, packet “mangling” techniques are used so that all packets originating from a network-computing device to its local LAN are spoofed to reflect the characteristics of the original connection. Thus Machine A is at all times of the impression that it communicating directly with Machine B, and vise versa. The existence of the network accelerators 14-1 and 14-2 completely unknown to Machines A or Machines B. In addition, both A and B are not aware that compression algorithms are employed.

[0041] In the preferred embodiment, the compression scheme used is a variation of LZ77 and Huffman coding compression algorithms. The original LZ77 algorithm is described in a paper by Ziv J., et al., “A Universal Algorithm for Sequential Data Compression,” IEEE Transactions on Information Theory, Vol. IT-23 (1979) pp. 337–343, although variants thereof can be used. The Huffman coding compression algorithm is described in “A Method for the Construction of Minimal Redundancy Codes,” Proceedings of the IRE, Vol. 40, (1952), pp. 1098–1101, although again, variants can be used In a preferred embodiment, compression occurs as follow Data is first compressed using an LZ77 algorithm. This algorithm uses a persistent compression dictionary associated with a persistent connection assigned to transfer the data. In the next step, a Huffman coding algorithm is then applied to the results the first step. If the results of the previous steps exceed the size of the original data, then the original data is sent as is.

[0042] While this invention has been particularly shown and described with references to preferred embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the scope of the invention encompassed by the appended claims.