Feed processing

Abstract

A data processing system comprising: a processing subsystem supporting a plurality of consumers, each consumer being arranged to process messages received into a corresponding receive queue; a network interface device supporting a virtual interface for each of the receive queues; and a hardware accelerator coupled to the processing subsystem by the network interface device and configured to parse one or more streams of data packets received from a network so as to, for each consumer: identify in the data packets messages having one or more of a set of characteristics associated with the consumer; and frame the identified messages in a new stream of data packets addressed to a network endpoint associated with the virtual interface of the consumer so as to cause said new stream of data packets to be delivered into the receive queue of the consumer.

Description

BACKGROUND

The invention relates to a data processing system and method for distributing messages received in streams of data packets to software consumers.

There are an increasing number of applications that require streams of network messages to be consumed and processed at very high data rates. For example, electronic financial exchanges typically provide several data feeds comprising messages carrying market parameters such as security buy/sell values or trade orders. Each of these feeds can have data rates that currently peak at over two million messages per second and present a considerable processing challenge to a computer system configured to receive the feeds, such as a bank trading system. Other examples of receivers that are required to process message streams at high data rates include servers hosting databases, file caches and webservers.

Generally, the size of each message in such systems will be relatively small compared to the size of data packets in which data is carried over a network to the receivers. Many messages are therefore packed into each data packet in order to maximise the efficiency of message transmission. This requires the receiver to parse the streams of data packets it receives in order to identify each message and pass it on to the appropriate consumer at the receiver.

Typically, the consumers running in software at a computer system configured to receive such data streams will each require only a subset of the messages contained in the streams. For example, trading software at a bank computer system would typically be configured to normalise and re-publish, or trade based on a limited number of securities.

So as to not overload the consumers with irrelevant messages, a dispatcher process is required to parse each data packet stream to identify the messages and then forward on to each consumer only those messages that are required. However, because of the sequential nature of parsing an individual data stream, the dispatcher for a given data-feed will be a process thread running at a single core and is therefore the limiting factor on the speed of the overall system. In other words, whilst the consumer software might be designed to efficiently distribute its processing across multiple cores of the system by executing threads in parallel, the dispatcher and hence the consumer software it serves remains limited by the speed of the core at which the dispatcher is running. This problem is more generally known as Amdahl's Law, which expresses that the speed of a parallelised computation is bound by the speed of the sequential portion of the computation. Even where the streams are provided as separate data-feeds (e.g., on separate IP multicast addresses) and therefore multiple dispatcher threads may be operating in parallel on some of the cores of the computer system, ultimately processing will be bottlenecked.

Furthermore, the bottleneck resulting from the use of a software dispatcher is generally aggravated by the fact that all of the messages of a data stream are provided to the host for handling at the dispatcher process. In most cases however, not all of the messages of a data stream are required by the consumers of the host and the dispatcher process therefore wastes resources and increases latency in the system by handling messages that are not wanted by the host.

There is therefore a need for an improved method for handling streams of data packets at a data processing system.

SUMMARY OF THE INVENTION

According to a first aspect of the present invention there is provided a data processing system comprising:

• • a processing subsystem supporting a plurality of consumers, each consumer being arranged to process messages received into a corresponding receive queue;
  • a network interface device supporting a virtual interface for each of the receive queues; and
  • a hardware accelerator coupled to the processing subsystem by the network interface device and configured to parse one or more streams of data packets received from a network so as to, for each consumer:
    • identify in the data packets messages having one or more of a set of characteristics associated with the consumer; and
    • frame the identified messages in a new stream of data packets addressed to a network endpoint associated with the virtual interface of the consumer so as to cause said new stream of data packets to be delivered into the receive queue of the consumer.

Preferably the hardware accelerator is configured to not forward said one or more streams of data packets received from the network.

Preferably the hardware accelerator is supported at the network interface device.

Preferably each new stream of data packets for a consumer is formed in accordance with a predetermined network protocol.

Suitably each new stream of data packets for a consumer is formed in accordance with the same network protocol in accordance with which received streams of data packets comprising messages for the new stream are formed.

Preferably the hardware accelerator is further configured to, on parsing each stream of data packets received from the network, identify control messages intended for all recipients of messages of that received data stream and include those control messages in each of the new streams of data packets directed to consumers that are recipients of messages from that received data stream. Alternatively the hardware accelerator is further configured to, on parsing each stream of data packets received from the network, identify control messages intended for all recipients of messages of that received data stream and include those control messages in a new streams of data packets directed to one or more predetermined consumers of the set of consumers that are recipients of messages from that received data stream.

Suitably at least one of the streams of data packets received from the network is a financial data feed from an electronic exchange and the control messages in that data stream intended for all recipients of messages of that received data stream are control messages relating to the operation of the electronic exchange and intended for all subscribers to that financial data feed.

Suitably each new stream of data packets for a consumer is formed in accordance with the same application layer protocols with which the received streams of data packets comprising messages for the new stream are formed.

Preferably the hardware accelerator is configured to duplicate messages that have characteristics associated with more than one consumer so as to provide those messages to each of those consumers in their respective new data streams.

Preferably the hardware accelerator is configured to do one of the following on receiving messages that do not have any of the characteristics associated with the consumers and are not control messages intended for all recipients of messages from the received data streams:

• • discard those messages;
  • deliver those messages in a data stream to one or more predetermined consumers; or
  • deliver each of those messages in a data stream to a consumer selected by means of a hash performed over one or more characteristics of that message.

Preferably the hardware accelerator is configured to, on receiving a stream of data packets that includes redundant forward error correction information, use the forward error correction information to attempt to recover any missing data packets and discard the forward error correction information prior to parsing the stream of data packets.

Preferably the hardware accelerator is configured to discard duplicate messages in a received stream of data packets.

Suitably the hardware accelerator is configured to, for a given consumer, convert messages identified in the received data streams for that consumer into a predefined data format prior to inclusion of the identified messages in the corresponding new data stream.

Suitably the hardware accelerator is configured to, prior to inclusion of the identified messages in a new data stream, process at least some of the identified messages by performing one or more of:

• • conversion of predefined data representations in the identified messages into host endian natural representations;
  • conversion of predefined numeric values in the identified messages into a binary numeric format;
  • decompression of the identified messages; and
  • discarding unwanted fields from the identified messages.

Preferably the hardware accelerator is configured to deliver data packets of a new data stream into its corresponding receive queue according to a reliable delivery mechanism.

The data processing system preferably further comprises, for each of the consumers, a protocol processing entity, the protocol processing entity configured to process the new stream of data packets so as to extract the identified messages for that respective consumer. Preferably the protocol processing entity is arranged to be supported at the same processing core as its respective consumer. Preferably each of the protocol processing entities is a user-level protocol processing entity. Preferably the receive queue of each respective consumer is held in memory mapped into the address space of that user-level protocol processing entity.

Suitably one or more of the consumers belong to a user-level application supported at an operating system of the processing subsystem, the application being configured to cause the operating system to load a consumer at each of a plurality of processing cores of the data processing system so as to distribute the consumers across the processing cores of the processing subsystem.

Preferably each consumer is configured to identify to the hardware accelerator at least some of the set of characteristics to be associated with that consumer.

Preferably the hardware accelerator or network interface device further comprises a data store holding for each of the consumers the set of characteristics associated with that consumer.

The data processing system preferably further comprises a software interface configured to, in response to an appropriately formatted request from a consumer that includes a representation of a set of characteristics identifying messages required by that consumer, cause that set of characteristics to be stored at the data store and associated with the consumer.

Suitably the software interface is configured to write the set of characteristics to the data store by means of a driver of the network interface device.

Suitably the software interface is an application programming interface.

Preferably the hardware accelerator is configured to update the sets of characteristics stored at the data store in respect of each consumer in dependence on information received in messages from the network.

Preferably the hardware accelerator is configured to identify in messages having one or more of a set of characteristics associated with a consumer associations between one or more characteristics of the current set and a new characteristic not currently associated with that consumer, the hardware accelerator being configured to store said new characteristic at the data store and associate the new characteristic with the consumer.

Suitably at least one of the consumers is a financial consumer configured to process messages relating to financial securities. Suitably the financial consumer is configured to identify a set of characteristics to the hardware accelerator that includes one or more of: security symbol, order ID, an identifier of the financial exchange at which the message originates, and a price or other financial parameter of a security or a market. Suitably the new characteristic is an order ID not yet known to the respective consumer and the one or more characteristics of the current set is a security symbol.

Suitably the processing subsystem and network interface device are coupled together by means of a data bus.

Preferably the network interface device is configured to write directly into the receive queues of the consumers by means of direct memory access.

Preferably the hardware accelerator is a reconfigurable logic device.

According to a second aspect of the present invention there is provided a method for distributing messages to consumers at a data processing system comprising a processing subsystem, a network interface device, and a hardware accelerator coupled to the processing subsystem by the network interface device, the processing subsystem supporting a plurality of consumers, each consumer being arranged to process messages received into a corresponding receive queue, and the method comprising:

• • a consumer requesting messages from one or more streams of data packets received at the hardware accelerator by indicating to the hardware accelerator a set of characteristics identifying those messages; and
  • on receiving the one or more streams of data packets from a network, the hardware accelerator parsing the streams of data packets and, for each consumer:
    • identifying in the data packets messages having one or more of the set of characteristics indicated to the network interface device by the consumer; and
    • framing the identified messages in a new stream of data packets addressed to a network endpoint associated with a virtual interface provided at the network interface device so as to cause said new stream of data packets to be delivered into the receive queue of the consumer.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will now be described by way of example with reference to the accompanying drawings, in which:

FIG. 1 is a schematic diagram of a data processing system configured in accordance with the present invention.

DETAILED DESCRIPTION OF THE DRAWINGS

The following description is presented to enable any person skilled in the art to make and use the invention, and is provided in the context of a particular application. Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

The present invention relates to data processing systems configured to support a plurality of message consumers in software. A data processing system configured in accordance with the present invention could be a computing device of any kind, such as a server or personal computer. The data processing system could include a plurality of processing cores each operable to support a consumer (although it should be noted that not all processing cores need support a consumer). Each processing core could be single core of a multi-core processor, or a distinct processor, with the processors being any kind of processing unit operable to support a software consumer.

A data processing system configured in accordance with the present invention is shown in FIG. 1. Data processing system 100 comprises a processing subsystem 101 coupled to a network 104 by means of a network interface device 102. Processing subsystem 101 is connected to network interface device (NIC) 102 by interconnect 103, which would typically be a data bus. The processing subsystem in this example comprises a plurality of processing cores 105 and a memory 106 (which could comprise any number, type or combination of memory modules). The network interface device comprises a hardware accelerator 109 coupled to the processing subsystem by network interface device 102. The hardware accelerator is preferably a reconfigurable logic device, such as an FGPA.

Processing subsystem 101 supports a software environment that includes consumers 107 for processing messages received in one or more streams of data packets received over network 104. The hardware accelerator is configured to parse data packets received from the network so as to identify the application layer messages required by the consumers. Each consumer is a software entity configured to receive and process messages from an associated receive queue. For example, a consumer could be any kind of program, application, process or thread supported at a processing core of the processing subsystem.

Each of the consumers receives data packets from NIC 102 into a corresponding receive queue 108 in memory 106. If the communication path between the accelerator and consumer is supported by a user-level networking stack of the processing subsystem it is preferable that the receive queue is locally addressable by the consumer such that the processing of messages occurs local to the receive queue (i.e. the receive queue is located in the address space of the consumer). If the communication path between the accelerator and consumer is supported by a kernel-level networking stack of a multi-core processing subsystem, it is preferable that the receive queue is at the same NUMA node as the consumer such that the processing of messages occurs local to that node of the system. The consumers could be part of a multi-threaded application configured to distribute its processing over a plurality of processing cores by arranging that its threads are supported at different processing cores of the system. Typically a memory controller for memory 106 and bus interface for data bus 103 would be integrated on the CPU dies supporting processing cores 105.

Data processing system 100 is arranged to receive one or more streams of data packets over network 104. On receiving the streams, accelerator 109 parses the data packets so as to identify the messages therein and extracts those messages having characteristics which indicate that one or more of the consumers 107 have requested to receive that message.

Preferably the network interface device (possibly the accelerator itself) includes a data store 110 identifying which messages each consumer has requested to receive. This can be achieved by arranging that the data store holds information representing the characteristics of messages requested by each consumer. Such characteristics could be any suitable identifiers in the received messages. For example, each message might include a header or be of a predetermined format such that the accelerator can be configured to compare the contents of predetermined fields or message parameters with those held in the data store. The characteristics could be provided to the NIC by or on behalf of the consumer, or could be updated by the accelerator itself in dependence on data received over the network.

For example, the data processing system could be a financial trading system at a bank receiving multiple data feeds from one or more electronic financial exchanges. In this example, the characteristics used to identify messages requested by a given consumer could be one or more of security symbol, order ID (e.g. for a buy/sell/offer/quote message), an identifier of the financial exchange at which the message originates, and a price or other financial parameter of a security or a market. Thus, a consumer at one processor could request all messages relating to a first set of securities by causing the corresponding set of security symbols and any related order IDs (e.g. relating to trades in that those securities by other parties) to be stored in the data store in respect of that consumer. Preferably the accelerator would be configured to update its data store in response to data received in messages from the network. For example, financial data feeds typically identify in a message to subscribers the underlying security to which each new order ID relates, and in subsequent messages regarding that order just the order ID would be used. The accelerator is preferably configured to store each new association so as to allow new messages that comprise only an order ID to be matched to the underlying security in which one or more consumers have registered an interest.

For messages received at the accelerator that do not match the characteristics associated with any consumer, the accelerator would preferably discard those messages, but could alternatively: (i) deliver those messages to a default consumer; (ii) deliver those messages to all consumers; (iii) spread those messages over a predetermined set of consumers (e.g. by hashing one or more characteristics and using the hash to distribute the messages evenly over the consumers.

Processing subsystem 101 preferably includes a software interface (not shown in the figures) such as an application programming interface (API) by means of which the consumers 107 or another entity on behalf of the consumers can submit requests for messages received at the network interface device. Each request includes an expression of the characteristics of the messages that the respective consumer wants to receive. The interface is configured to receive the message requests from the consumers or other software entities managing the consumers and/or forming part of a greater application of which the consumers form a part. In response to the requests, the interface causes appropriate information to be written to the data store to express the set of message characteristics requested by the consumer.

The interface would typically effect the writing of requested message characteristics to the data store by means of a driver of the network interface device or accelerator, or another entity permitted to access the data store at the network interface device/accelerator. In this manner, the messages required by each consumer can be registered at the network interface device/accelerator.

Typically the processing subsystem would comprise an operating system supporting a user-level application environment. The consumers could be provided at the operating system but would preferably be at user level. The software architecture of the processing subsystem could take any suitable form: for example, the subsystem could comprise one or more virtualised operating systems supported atop a hypervisor or other similar software layer, and the consumers could be supported by one or more different virtualised operating systems. For maximum performance, a user-level consumer could be provided with a user-level protocol stack that is interfaced directly to a virtual network interface of the NIC.

On extracting the requested messages from the streams of data packets received at the NIC, the accelerator is configured to form new data packets streams, one for each of the consumers supported at processing cores 105. The new data packet streams are preferably formed in memory at accelerator 109. The memory could be the same memory that holds data store 110. The “new” data packet streams are new in the sense that the streams originate at the hardware accelerator; once established, the new data packet streams are preferably re-used to communicate data packets formed at the accelerator to each respective consumer. The data packets for each of the consumers comprise the messages requested by that consumer; thus, a given message could be duplicated across more than one of the new data streams. The headers for the new streams would be generated according to the delivery requirements of the consumers. For example, each generated new data stream could have a unique new IP destination address so as to cause each data stream to be delivered to the correct endpoint/receive queue. The data packets of the streams received at the accelerator are preferably not passed onto the processing subsystem and only the new streams of data packets are written into the receive queues of the host. Most preferably the hardware accelerator is configured to not forward onto any network endpoints said one or more streams of data packets received from the network, said network endpoints including endpoints of the processing subsystem.

Any control messages relevant to the consumers of a given data stream received at the NIC are preferably passed onto those consumers that have associated characteristics indicating one or more messages in that received data stream. For example, if a received data stream is a financial data feed from an electronic exchange, that data feed would typically include control messages intended for all consumers receiving messages from that feed, such as control messages indicating the start and end of trading each day so as to inform subscribers to the feed when trading at the exchange is to start and stop. Alternatively, one or more predetermined consumers could be arranged to receive such control messages—this can be useful if a plurality of consumers of the system belong to a single application, with one of those consumers being configured to receive control messages relating to the data streams the application has requested to receive.

Accelerator 109 is configured to encapsulate the messages requested by each consumer in data packets configured according to any suitable network protocol. Each consumer could require streams of data packets according to a different protocol but preferably the network protocol would be the same for all new data streams. The protocols by which data packets of the new streams comprising a given set of messages are formed could differ from the protocols by which the data packets comprising those messages are received at the NIC. Thus, data packet streams of different network protocols that are received at the NIC could be normalised for each consumer into new data packet streams according to a predetermined network protocol. Thus, in the example that the data processing system 100 is a trading system, the NIC could receive two streams: a first stream of data packets according to the UDP protocol and a second stream of data packets according to the TCP protocol, with new streams created at the accelerator and comprising messages from both incoming packet streams being, for example, streams of UDP data packets.

Accelerator 109 could be further configured to convert received messages into a predefined message format prior to including those messages in a data stream for a consumer. This can have performance advantages because the accelerator can be configured to perform processing that can be done more efficiently in hardware than in software (e.g. decompression of compressed messages) and it can ensure that all messages are received at the consumer in a common format (e.g. all messages are FIX messages). Examples of the processing that could be performed by the accelerator in order to convert received messages into the correct format for a given consumer include: conversion of predefined data representations into host endian natural representations; conversion of predefined numeric values into a binary numeric format; and decompression of the received messages that are identified as being for inclusion in a data packets of a new data stream for the given consumer. Performing decompression can be useful when the received streams of data packets are financial data feeds from an electronic exchange and the messages are compressed according to the FIX/FAST protocol. Performing decompression of compressed FIX/FAST messages at the accelerator reduces latency and improves performance of the trading system.

Preferably the data packet streams created at the accelerator for the consumers are UDP protocols or another lightweight, preferably connectionless, protocol. Since the data packets need travel only from the NIC up to a receive queue at processing subsystem 101, the data packets could be formed without any check data (i.e. the step of forming packet checksums could be omitted). Preferably the data packets are delivered to the processing subsystem according to a reliable data protocol.

The data packets of the new streams generated at the accelerator are directed to network endpoints associated with the appropriate consumer such that the messages requested by a process are delivered in a new stream of data packets into the receive queue of that consumer. This is achieved by arranging that the NIC provides a virtual interface for each consumers receive queue. By directing the data packet streams to predetermined endpoints (which could be defined in data store 110 but preferably are held in state at accelerator 109), the accelerator can cause each new stream to be delivered into the appropriate receive queue. By arranging that a consumer and its receive queue are local to one another (e.g. at the same NUMA node or in, or mapped into, the same user level address space of the same virtualised operating system at the host), each new stream of data packets is delivered into a memory location which is local to the corresponding consumers. This allows the new streams of data packets to be consumed without context switching and without any forwarding of the new data packet streams.

Preferably a hardware controller 111 of the network interface device is configured to provide the virtual interfaces and handle the transmission of data packets over interconnect 103 into the receive queues 107 of the host processing subsystem in the conventional manner. The transmission of data packets over interconnect 103 is preferably performed by means of direct memory access (DMA) transfers into a memory supporting the receive queues.

As the data packets of each new stream are received into the respective receive queues, those data packets must be processed according to the relevant network protocols in order to extract the messages therein. It is advantageous if a protocol processing entity (such as a transport library) is provided at each consumer and configured to perform the protocol processing of data packets received at the receive queue of the consumer supported at that core. This ensures that both the protocol processing of a given data packet and the processing of the messages therein by the relevant consumer takes place on the same core, and avoids the costly overheads such as cache bouncing that occur when processing moves between cores. The protocol processing entities could be configured to post the messages into an intermediate queue accessible to the consumers.

In the case that the consumers are user level consumers, it is advantageous if the protocol processing entities provided at the cores are user level entities and the receive queues are mapped into the address space of the user level entities so as to avoid context switches between the kernel and user level when data packets are consumed from the receive queues.

In alternative embodiments, NIC 102 could be configured to perform at least some protocol processing such that the messages contained within the data packets of each stream are extracted at the NIC (e.g. at an offload engine of controller 111) and posted into the corresponding receive queues for consumption directly by the consumers.

In accordance with the present invention, the accelerator receives one or more streams of data packets from a network and pushes out a customised stream of data packets to each of the consumers. In this manner, the accelerator distributes in hardware the processing of the messages received in the data packets streams across the consumers of the system without requiring a software dispatcher to parse and split out the received messages. By arranging that the consumers are distributed over the processing cores of the system the present invention therefore allows the burden of message processing to be evenly shared over the cores of the system.

In certain embodiments, the NIC might receive one or more redundant data streams such that if any data packets are missing from one stream there is a chance the data packet will be available from a redundant stream. In such embodiments, the accelerator is preferably configured to aggregate the streams into a single stream so as to recover any missing data packets and discard any duplicate messages prior to parsing the single stream of data packets.

In certain embodiments of the present invention, the NIC might receive a stream which contains redundant forward error correction (FEC) information. In such embodiments, the accelerator is preferably configured to use the FEC information to recover any missing data packets and discard redundant information prior to parsing the stream of data packets.

The NIC and accelerator described herein need not be provided at the same device and could be, for example, separate peripheral boards of data processing system 100 connected together by means of an interconnect (e.g. bus 103) or a dedicated link.

The applicant hereby discloses in isolation each individual feature described herein and any combination of two or more such features, to the extent that such features or combinations are capable of being carried out based on the present specification as a whole in the light of the common general knowledge of a person skilled in the art, irrespective of whether such features or combinations of features solve any problems disclosed herein, and without limitation to the scope of the claims. The applicant indicates that aspects of the present invention may consist of any such individual feature or combination of features. In view of the foregoing description it will be evident to a person skilled in the art that various modifications may be made within the scope of the invention.

Claims

1. A network interface device comprising: circuitry configured to support a virtual interface for each of a plurality of receive queues associated with a plurality of application layer consumers comprised within a processing subsystem;a hardware accelerator configured to parse one or more streams of data packets received from a network so as to: identify application layer messages within the received data packets and extract application layer messages having one or more of a respective set of characteristics requested by a respective application layer consumer;duplicate application layer messages that have one or more of the respective set of characteristics associated with more than one application layer consumer so as to provide those application layer messages to each of those application layer consumers; andframe the identified application layer messages in a new stream of data packets for each application layer consumer, by encapsulating the identified application layer messages with headers identifying each respective application layer consumer into data packets, such that the data packets are addressed to a network endpoint associated with the virtual interface of the respective application layer consumer so as to cause said new stream of data packets to be delivered into the receive queue of the respective application layer consumer.

2. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to not forward said one or more streams of data packets received from the network.

3. A network interface device as claimed in claim 1, wherein each new stream of data packets for an application layer consumer is formed in accordance with a predetermined network protocol.

4. A network interface device as claimed in claim 1, wherein each new stream of data packets for an application layer consumer is formed in accordance with the same network protocol in accordance with which received streams of data packets comprising application layer messages for the new stream are formed.

5. A network interface device as claimed in claim 1, wherein the hardware accelerator is further configured to, on parsing each stream of data packets received from the network, identify control messages intended for all recipients of application layer messages of that received data stream and include those control messages in each of the new streams of data packets directed to application layer consumers that are recipients of application layer messages from that received data stream.

6. A network interface device as claimed in claim 5, wherein at least one of the streams of data packets received from the network is a financial data feed from an electronic exchange and the control messages in that data stream intended for all recipients of application layer messages of that received data stream are control messages relating to the operation of the electronic exchange and intended for all subscribers to that financial data feed.

7. A network interface device as claimed in claim 1, wherein the hardware accelerator is further configured to, on parsing each stream of data packets received from the network, identify control messages intended for all recipients of application layer messages of that received data stream and include those control messages in a new streams of data packets directed to one or more predetermined application layer consumers of the set of application layer consumers that are recipients of application layer messages from that received data stream.

8. A network interface device as claimed in claim 1, wherein each new stream of data packets for an application layer consumer is formed in accordance with the same application layer protocols with which the received streams of data packets comprising application layer messages for the new stream are formed.

9. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to do one of the following on receiving application layer messages that do not have any of the characteristics associated with the application layer consumers and are not control messages intended for all recipients of application layer messages from the received data streams: discard those application layer messages;deliver those application layer messages in a data stream to one or more predetermined application layer consumers; ordeliver each of those application layer messages in a data stream to an application layer consumer selected by means of a hash performed over one or more characteristics of that application layer message.

10. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to, on receiving a stream of data packets that includes redundant forward error correction information, use the forward error correction information to attempt to recover any missing data packets and discard the forward error correction information prior to parsing the stream of data packets.

11. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to discard duplicate application layer messages in a received stream of data packets.

12. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to, for a given application layer consumer, convert application layer messages identified in the received data streams for that application layer consumer into a predefined data format prior to inclusion of the identified application layer messages in the corresponding new data stream.

13. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to, prior to inclusion of the identified application layer messages in a new data stream, process at least some of the identified application layer messages by performing one or more of: conversion of predefined data representations in the identified application layer messages into host endian natural representations;conversion of predefined numeric values in the identified application layer messages into a binary numeric format;decompression of the identified application layer messages; anddiscarding unwanted fields from the identified application layer messages.

14. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to deliver data packets of a new data stream into its corresponding receive queue according to a reliable delivery mechanism.

15. A network interface device as claimed in claim 1, wherein each application layer consumer is configured to identify to the hardware accelerator at least some of the set of characteristics to be associated with that application layer consumer.

16. A network interface device as claimed in claim 1, wherein the hardware accelerator or network interface device further comprises a data store holding for each of the application layer consumers the set of characteristics associated with that application layer consumer.

17. A network interface device as claimed in claim 16, further comprising a software interface configured to, in response to an appropriately formatted request from a application layer consumer that includes a representation of a set of characteristics identifying application layer messages required by that application layer consumer, cause that set of characteristics to be stored at the data store and associated with the application layer consumer.

18. A network interface device as claimed in claim 17, wherein the software interface is configured to write the set of characteristics to the data store by means of a driver of the network interface device.

19. A network interface device as claimed in claim 17, wherein the software interface is an application programming interface.

20. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to update the sets of characteristics stored at the data store in respect of each application layer consumer in dependence on information received in application layer messages from the network.

21. A network interface device as claimed in claim 1, wherein the hardware accelerator is configured to identify in application layer messages having one or more of a set of characteristics associated with an application layer consumer associations between one or more characteristics of the current set and a new characteristic not currently associated with that application layer consumer, the hardware accelerator being configured to store said new characteristic at the data store and associate the new characteristic with the application layer consumer.

22. A network interface device as claimed in claim 21, wherein at least one of the application layer consumers is a financial application layer consumer configured to process application layer messages relating to financial securities.

23. A network interface device as claimed in claim 22, wherein the financial application layer consumer is configured to identify a set of characteristics to the hardware accelerator that includes one or more of: security symbol, order ID, an identifier of the financial exchange at which the application layer message originates, and a price or other financial parameter of a security or a market.

24. A network interface device as claimed in claim 22, wherein the new characteristic is an order ID not yet known to the respective application layer consumer and the one or more characteristics of the current set is a security symbol.

25. A network interface device as claimed in claim 1, wherein the network interface device is coupled to the processing subsystem by means of a data bus.

26. A network interface device as claimed in claim 1, wherein the network interface device is configured to write directly into the receive queues of the application layer consumers by means of direct memory access.

27. A network interface device as claimed in claim 1, wherein the hardware accelerator is a reconfigurable logic device.

28. A method for distributing application layer messages to application layer consumers, the method comprising: receiving a request, from a plurality of application layer consumers, for application layer messages from one or more streams of data packets received at the hardware accelerator by indicating to the hardware accelerator a respective set of characteristics identifying those application layer messages; andon receiving the one or more streams of data packets from a network, the hardware accelerator parsing the streams of data packets and: identifying application layer messages within the received data packets and extracting those application layer messages having one or more of the respective set of characteristics requested by a respective application layer consumer;duplicating the application layer messages that have one or more of the respective set of characteristics associated with more than one application layer consumer so as to provide those application layer messages to each of those application layer consumers; andframing the identified application layer messages in a new stream of data packets for each application layer consumer, by encapsulating the identified application layer messages with headers identifying each respective application layer consumer into data packets, such that the data packets are addressed to a network endpoint associated with a virtual interface of the respective application layer consumer provided at the network interface device so as to cause said new stream of data packets to be delivered into the receive queue of the respective application layer consumer.

29. A data processing system comprising: a processing subsystem supporting a plurality of application layer consumers, each application layer consumer being arranged to process application layer messages matching a respective set of characteristics requested by that application layer consumer received into a respective receive queue associated with that application layer consumer;a network interface device supporting a virtual interface for each of the receive queues and comprising a hardware accelerator, said hardware accelerator configured to parse one or more streams of data packets received from a network so as to: identify application layer financial data messages within the data packets and extract those application layer messages having one or more of the respective set of financial characteristics requested by a respective application layer consumer;duplicate the application layer financial messages that have one or more of the respective set of financial characteristics associated with more than one application layer consumer so as to provide those application layer messages to each of those application layer consumers; andframe the identified application layer messages in a new stream of data packets for each application layer consumer, by encapsulating the identified application layer messages with headers identifying each respective application layer consumer into data packets, such that the data packets are addressed to a network endpoint associated with the virtual interface of the respective application layer consumer so as to cause said new stream of data packets to be delivered into the receive queue of the respective application layer consumer.

30. A network interface device comprising: circuitry configured to support a virtual interface for each of a plurality of receive queues, each receive queue being associated with a respective application layer consumer supported by a processing subsystem;a hardware accelerator configured to parse one or more streams of data packets received from a network so as to: identify application layer financial data messages within the data packets and extract those application layer messages having one or more of a respective set of financial characteristics requested by a respective application layer consumer;duplicate the application layer financial messages that have one or more of the respective set of financial characteristics associated with more than one application layer consumer so as to provide those application layer messages to each of those application layer consumers; andframe the identified application layer messages in a new stream of data packets for each application layer consumers, by encapsulating the identified application layer messages with headers identifying each respective application layer consumer into data packets, such that the data packets are addressed to a network endpoint associated with the virtual interface of the respective application layer consumer so as to cause said new stream of data packets to be delivered into the receive queue of the respective application layer consumer.