Method for interconnecting nodes of a real-time parallel computer

p0001The present invention relates to a method for interconnecting the nodes of a parallel real-time computer in which is used as an interconnection system, a local network switching virtual channels, each virtual channel being associated with a determined bandwidth, of to ensure control of end to end latency of transmission of a message to be transmitted in real time, in a chain comprising a transmitter node with a main processor and an interconnect interface adapter, a receiver node with a main processor and an interconnect interface adapter, and at least one switch connecting said interface adapters. It also relates to an interconnection interface adapter for the implementation of this method.

p0002The present invention relates to the field of parallel computing systems (multi-node), and particularly to those used in real-time distributed environments. The process according to the invention is applicable whenever a parallel computing system requires high bandwidth for data transfers between nodes, along with a deterministic and low latency for real-time communications.

p0003It is known that in a system having a plurality of computing nodes connected by an interconnection system, it is necessary to exchange information, type command or data via messages conveyed by the interconnect system.

p0004One method is to connect the compute nodes by a specific communication network, local, obtaining a regular network topology, fixed type O, matrix, etc ... These interconnect systems are used in type machines MPP (Massively Parallel Processors).

p0005This method generally used to connect a relatively large number of compute nodes with good performance in terms of bandwidth. The major drawback of this method is that the latency of message transmission, from end to end, is very poorly controlled because the data type information is transferred by large blocks to use the network bandwidth and messages control are not differentiated.

p0006In addition, these interconnection systems do not allow a real geographical distribution of computing nodes. Thus, the acquisition of remote data can be done by the input / output interfaces, prohibiting local processing of data, or a local network connected to one of the computer nodes, which increases latency transfer to these acquisitions.

p0007A second approach, used today, is the computer network layout, this method being known as the "cluster" (cluster). In this method, the priorities of the transferred data is not transmitted or processed by the network, which prohibited to control the overall system behavior. In addition, the operation of the equipment requires software support to control the interconnection network and ensure reliable transport of information, which increases latency transmissions a significant factor and can not control the latency necessary a real-time operation.

p0008The present invention therefore a main object to remedy these drawbacks and to do that, it relates to a process for obtaining a deterministic and low latency between nodes of a real-time parallel computer using the switch and physical interface of a LAN switch, as an interconnection system.

p0009This method is essentially characterized in that it comprises the steps of:<ul><li>to allocate, during initialization of an application in real time on the main processor of the originating node, the virtual channel bandwidth for intensive data to be transmitted and virtual channel bandwidth reserved for data to be transmitted in real time ;</li><li>to assign a priority level to each message to be transmitted in real time;</li><li>segmenting data intensive and real-time data into data blocks, adding to each real-time data block information corresponding priority message; </li><li>to pass on the adapter of the transmitter node interconnect interface data blocks corresponding to the intensive data at a rate corresponding to the available bandwidth of the virtual channel in question and the data blocks corresponding to the data to be transmitted in real time at a rate corresponding to the reserved bandwidth of the virtual channel, while respecting the priority level assigned to each data block;</li><li>reassembling the data blocks in the adapter of the receiving node interconnect interface to reconstitute the messages; and</li><li>to transmit them to the main processor of the receiver node in accordance with the priority level of said messages.</li></ul>

p0010Thus, the real time aspects are taken into account throughout. Among others, the determinism of the latency is achieved by a data message of differentiation and those carrying control of the software execution from the node sending the message, its interface with the communication network, the network itself, the interface with the network of the receiving node and the receiving node. This process remains valid when the transferred data is voice, images or video.

p0011Finally, the proposed approach allows the same interconnection for internal and external communications with an adaptation of the flow as needed. Integration in a communications network is easy while providing real-time behavior both internally and externally.

p0012It is also noted that the method according to the invention allows to certain nodes of the parallel computer at remote locations, particularly the data acquisition nodes, depending on the capabilities of the local communication network.

p0013According to the invention, an interconnection interface adapter for the implementation of this method is essentially characterized in that it comprises: <u>the transmitting end</u><ul><li>means to store real-time messages, created by the main processor of the sending node and previously assigned a priority level, in message transmission queues considering each virtual channel used and the priority level for the within that channel,</li><li>arbitration means for selecting the following message to be transmitted and the corresponding virtual channel, based on the priority level of each message,</li><li>means for segmenting the message to be transmitted in blocks of data containing information of priority of the corresponding message,</li><li>data block formatting means, and</li><li>means for transmitting said data blocks on the local switching network comprising at least one switch,</li></ul><u>and the receiving side</u><ul><li>data blocks of the receiving means having passed through said LAN switch,</li><li>ways to reassemble the message from the received data blocks, and</li><li>means for storing messages reconstituted in reception tails taking account of each virtual channel used and the priority level of the message within this channel before transmitting them to the main processor of the receiver node.</li></ul>

p0014An embodiment of the invention is described below by way of example, with reference to the accompanying drawings in which: <ul><li>1 shows very diagrammatically the interconnection of nodes in a parallel computer by means of a LAN switch; and</li><li>2 illustrates schematically the different steps of transmitting a message between two nodes of Figure 1, according to the method of the invention.</li></ul>

p0015By first referring to Figure 1, there is shown a plurality of nodes A, B, ... X, Y of a parallel real-time calculator, interconnected by means of a switched network comprising a local network 1 switches and physical media such as 2_AT, 2_B... 2_X and 2_Y.

p0016Each of the nodes A, B, ... X, Y includes a main processor, respectively 3_AT, 3_B... 3_X, 3_Y accessing the local switched network by means of an interconnection interface adapter, respectively 4_AT, 4_B... 4_X, 4_Y, Via an input / output bus, respectively 5_AT, 5_B... 5_X, 5_Y. According to the invention, each adapter 4_AT, 4_B... 4_X, 4_Y provides support for real-time communications, in addition to all standard characteristics to the local network switching.

p0017Note that in the particular embodiment of the invention described here by way of example, the LAN switching is used for the ATM type (Asynchronous Transfer Mode). In such an ATM network, the communication between nodes is by means of virtual channels using the physical links 2_AT, 2_B, ... 2_X, 2_Y and which are associated with "QoS" representing collateral flow parameter data or bandwidth.

p0018According to the invention, the constant flow QoS called CBR is used to transfer data with real-time characteristics and quality of available broadband service called ABR is used to transfer data intensive, allowing 1 switches to use priority mechanisms to differentiate these two types of data and to ensure that the priority processing method for real-time data is compatible with the standard operation of said switches.

p0019We will now describe the real-time data transfer process between two nodes of the parallel computer, for example between the node A functioning as a transmitter and node B operates as a receiver, with particular reference to Figure 2 .

p0020When a real-time application running on the processor 3_AT Node A creates a message to transmit real-time data to the real-time application running on the processor 3_B Node B, it transmits in the header of the message the current priority assigned to said message. This is ensured by the operating system or by the application itself.

p0021The different messages thus created are then queued into means for queuing transmission 6_AT. According to the invention, a message queue is used for each virtual channel CBR VC respectively₁VC ..._iVC ..._at, And for each priority level within this virtual channel, respectively P₁... P_i... P_not.

p0022Note here that the ABR virtual channels that do not promote real-time data, but only data intensive, do not need to be organized by priority. Intensive data are transmitted between different nodes of the parallel computer in the same way that real-time data according to their own qualities of service without priority consideration.

p0023The selection of the message to be transmitted is through a means of arbitration 7_AT. These arbitration means first selects the virtual channel in accordance with the QoS parameter, as specified in the operation of the local network switching considered here the ATM network, then, in accordance with the invention selects in the tail transmission of this virtual channel the message having the highest priority level.

p0024If for example it is assumed that the virtual channel VC_i has been selected, the arbitration means 7_AT will then select the VC channel_i non empty message queue having the highest priority, for example the tail P_i. The selected message will be the message first entered the queue P_i.

p0025Note that each message whose transmission was interrupted in favor of a higher priority message references stored in the tail of transmission of message virtual channel allocated to it, until complete transmission of the message. Assuming that there was a transmission of the message being for this virtual channel VC_iLess priority than the message of the queue P_iThe references of this message is stored in the corresponding priority queue and an address indicating which part of the message is to be transmitted. This message will again be selected by means of arbitration 7_AT after all higher priority queues as his VC for this virtual channel_i depleted.

p0026The message selected by the arbitration means 7_AT is then segmented into data blocks or cells by the segmentation means 8_ATSo as to ensure the transfer of the message cell by cell. According to the invention, and when it is a message to be transmitted in real time, the payload of the cell or block of data is reduced to insert the priority information of the message, as given by the application and used for queuing transmission 6_AT and arbitration 7_AT.

p0027The cell is then formatted 9_ATIn accordance with the specification of the local network switching considered here the ATM network, so as to be correctly transmitted by the transmission means 10_AT the physical bonds such as 2_AT and treated properly by the switches of the interconnection network 1 with the same quality service.

p0028It should also be noted that the cells or blocks of data corresponding to the intensive data is transmitted to the adapter 4_AT at a rate corresponding to the bandwidth of the considered ABR virtual channel, while the cells corresponding to the real time data is transmitted on said adapter to a rate corresponding to the reserved bandwidth of the considered CBR virtual channel.

p0029After passing through the switch array 1, the cell enters the adapter 4_B, ... Of the receiving node B by the physical link 2_B. When the means of receipt 11B receive a cell, they separate the payload and control information, including priority information. This information is then used to reassemble the message in ways reassembly 12_B. According to the invention, a message queue is used for each virtual channel CBR and for each priority level within this virtual channel. All messages can be simultaneously reassembled regardless interleaving cell order between messages.

p0030When the full message is reassembled, the reference is placed in a reception queue 13_B organized by priority, according to the message priority. This priority is then taken into account by the real-time application running on the processor 3_B the receiver node B. It will be noted a determined moment, the receiving means of queuing 13_B can contain multiple messages from different nodes and with different priorities.

p0031Of course, all the adapters 4_AT, 4_B... 4_X, 4_Y different nodes are identical in construction and contain both the means necessary for the transmission and reception of messages.

p0032We see finally that the process according to the invention interconnect the nodes of a parallel computer can transmit at each moment of the cell with the highest priority message respecting the quality of service and priority of each message, data from applications running on different processors. This method thus provides a deterministic, low-latency transmission of real-time messages between the different nodes of the computer.

p0033Note also that such an interconnection method allows, due to the use of a LAN to certain nodes of the parallel computer to remote location, in particular the data acquisition nodes, providing great flexibility to the system.