Abstract: The invention relates to a method for generating a request, from a formal language instruction defining a set of ports of an interconnection network, said request including an addressing command for each one of the ports defined in the instruction, said method including the following steps: Receiving, by a communication module, a formal language instruction defining a set of ports, Processing, by a processing module, the formal language instruction so as to generate a set of numbers encoded on at least one byte, each number including position bits, each one of the position bits allowing to identify a port and at least one authorization bit, the at least one authorization bit allowing to define access rights on the ports, and Encoding, by an encoding module, the set of numbers so as to generate the request including the addressing command.

Abstract: Processes pending execution in respective compute nodes (N1-Nn) of a cluster of a distributed HPC computer can communicate with one another by message exchange through an interconnected fabric. To exchange messages between the processes a method is proposed to identify the physical cards associated with the compute nodes directly from the hostname of the compute nodes as they are used in the user program. This direct identification is made from at least one mapping table (MT1) associating bijectively the hostname of each compute node of the cluster with the unique logical address of the associated physical card. This mapping table is kept in a component of the computer responsible for managing the fabric, i.e. the fabric manager (FM1). Various implementations enable scaling when accomplishing the method of message exchange between processes.

Abstract: The invention relates to a method for generating a request, from a formal language instruction defining a set of ports of an interconnection network, said request including an addressing command for each one of the ports defined in the instruction, said method including the following steps: Receiving, by a communication module, a formal language instruction defining a set of ports, Processing, by a processing module, the formal language instruction so as to generate a set of numbers encoded on at least one byte, each number including position bits, each one of the position bits allowing to identify a port and at least one authorization bit, the at least one authorization bit allowing to define access rights on the ports, and Encoding, by an encoding module, the set of numbers so as to generate the request including the addressing command.

Abstract: The invention relates to a method for generating a request, from a formal language instruction defining a set of ports of an interconnection network, said request including an addressing command for each one of the ports defined in the instruction, said method including the following steps: Receiving, by a communication module, a formal language instruction defining a set of ports, Processing, by a processing module, the formal language instruction so as to generate a set of numbers encoded on at least one byte, each number including position bits, each one of the position bits allowing to identify a port and at least one authorization bit, the at least one authorization bit allowing to define access rights on the ports, and Encoding, by an encoding module, the set of numbers so as to generate the request including the addressing command.

Abstract: The invention relates to a device for acquiring values of counters associated with a computational task implemented on an interconnection network, said interconnection network including a plurality of nodes interconnected via switches, said nodes and switches including a plurality of ports associated with counters, the values of said counters being stored over time, said computational task involving a subset of ports associated with subsets of nodes and switches, and said device comprising: a routing acquisition module, able to acquire a routing table including all the connections of the interconnection network, a module for acquiring involved nodes, able to acquire a list of the nodes involved in the computational task, a port extraction module, able to extract a list of the ports involved in the computational task from the list of involved nodes and the routing table, and a counter extraction module, able to extract the values of the counters of the ports involved in the computational task from the list

Abstract: A method of remote monitoring and remote control of a cluster including nodes connected to a communication network of a first type, a relay node of the nodes including first and second network interfaces in accordance with first and second types of communication network respectively, the first and second types being distinct, some steps of the monitoring and control method being implemented in a remote computer linked to the relay node by a communication network of the second type, the method including receiving a packet via the first network interface; encapsulating the received data packet in a data frame in accordance with a protocol of the communication network of the second type; sending the data frame to the remote computer via the second network interface, the receiving of the data packet, the encapsulating of the data packet and the sending of the data frame being implemented in the relay node.

Abstract: The invention relates to a method for generating a request, from a formal language instruction defining a set of ports of an interconnection network, said request including an addressing command for each one of the ports defined in the instruction, said method including the following steps: Receiving, by a communication module, a formal language instruction defining a set of ports, Processing, by a processing module, the formal language instruction so as to generate a set of numbers encoded on at least one byte, each number including position bits, each one of the position bits allowing to identify a port and at least one authorization bit, the at least one authorization bit allowing to define access rights on the ports, and Encoding, by an encoding module, the set of numbers so as to generate the request including the addressing command.

Abstract: The invention relates to a device for acquiring values of counters associated with a computational task implemented on an interconnection network, said interconnection network including a plurality of nodes interconnected via switches, said nodes and switches including a plurality of ports associated with counters, the values of said counters being stored over time, said computational task involving a subset of ports associated with subsets of nodes and switches, and said device comprising: a routing acquisition module, able to acquire a routing table including all the connections of the interconnection network, a module for acquiring involved nodes, able to acquire a list of the nodes involved in the computational task, a port extraction module, able to extract a list of the ports involved in the computational task from the list of involved nodes and the routing table, and a counter extraction module, able to extract the values of the counters of the ports involved in the computational task from the list

Abstract: The method relates to the monitoring of at least one routing parameter for a cluster including nodes and switches, static communication links connecting nodes and switches. Each switch includes several output ports. After having selected at least one switch, a number of routes per port is calculated for each port of each switch selected, routes being defined during a routing step for each connecting one node to another. A mean number of routes per port is then calculated for the at least one selected switch. Each number of routes per port calculated is then compared with the mean number of routes per port calculated and, in response to this comparison, a potential imbalance of routing of the cluster is notified.

Abstract: A system comprising a peripheral having a timing mechanism and a node, one of which comprises a real memory space and the other a corresponding virtual memory space, is disclosed. On receiving a timing command in the real memory space, comprising references to an event and time, an entry comprising data relative to the event and time references is created in a monitoring queue of the peripheral. A current point in time is then compared, in the peripheral, to a scheduled point in time linked to an item of data relative to a time reference stored in the monitoring queue. In response, if the current point in time is after the scheduled point in time, an item of data relative to a reference linked to the item of data relative to a time reference stored in the monitoring queue is stored in the real memory space.

Abstract: Processes pending execution in respective compute nodes (N1-Nn) of a cluster of a distributed HPC computer can communicate with one another by message exchange through an interconnected fabric. To exchange messages between the processes a method is proposed to identify the physical cards associated with the compute nodes directly from the hostname of the compute nodes as they are used in the user program. This direct identification is made from at least one mapping table (MT1) associating bijectively the hostname of each compute node of the cluster with the unique logical address of the associated physical card. This mapping table is kept in a component of the computer responsible for managing the fabric, i.e. the fabric manager (FM1). Various implementations enable scaling when accomplishing the method of message exchange between processes.

Abstract: A method of remote monitoring and remote control of a cluster including nodes connected to a communication network of a first type, a relay node of the nodes including first and second network interfaces in accordance with first and second types of communication network respectively, the first and second types being distinct, some steps of the monitoring and control method being implemented in a remote computer linked to the relay node by a communication network of the second type, the method including receiving a packet via the first network interface; encapsulating the received data packet in a data frame in accordance with a protocol of the communication network of the second type; sending the data frame to the remote computer via the second network interface, the receiving of the data packet, the encapsulating of the data packet and the sending of the data frame being implemented in the relay node.

Abstract: The method relates to the monitoring of at least one routing parameter for a cluster including nodes and switches, static communication links connecting nodes and switches. Each switch includes several output ports. After having selected at least one switch, a number of routes per port is calculated for each port of each switch selected, routes being defined during a routing step for each connecting one node to another. A mean number of routes per port is then calculated for the at least one selected switch. Each number of routes per port calculated is then compared with the mean number of routes per port calculated and, in response to this comparison, a potential imbalance of routing of the cluster is notified.

Abstract: A method for allocating, in order to carry out a calculation, at least one first resource of a plurality of interconnected resources, the first resource being connected to a first port of a switch, the method including acquiring a first weight of the first resource, the weight corresponding to the number of resources of the plurality of resources connected to the first port of the switch.

Abstract: The invention relates in particular to the optimization of routing in a cluster comprising a plurality of nodes and static communication links connecting nodes of the plurality of nodes, said routing being based on load levels associated with the communication links. In order to establish a connection between at least two nodes of the cluster that have been identified (505), at least one route is determined (510) that connects the identified nodes according to the communication links, said route being determined according to the nodes identified, communication links and at least one load level associated with each communication link. A determined route is selected. Subsequently, a value of weight associated with the selected route is estimated (520) and a load level associated with each communication link of the selected route is incremented (525).

Abstract: The invention relates to the pseudo-dynamic routing in a cluster comprising nodes, of the static communication links connecting these nodes. The routing is based on load levels associated with the links. After having received (635) a list of node identifiers, these identifiers designating a set of nodes allotted to the execution of a task, a weight having a first predetermined value is assigned (655) to a pair formed of an identifier of a first and of a second node of said set of nodes, the first and second nodes being distinct, a weight having a second predetermined value, distinct from the first value, being assigned by default to the formed pairs of identifiers of nodes allotted to different tasks. The cluster is then routed, the routing comprising the selecting (610) of a route between the first and second nodes and the incrementing (625) of a load level associated with each link comprising the selected route of the first value.

Abstract: The invention relates in particular to a computer system including peripheral devices (600) and at least one switch (605) connected to each device. A first device includes a means for initiating a control of direct access to memory areas, each one of which is associated with a separate element of the system. The switch includes a means for transmitting at least a portion of the control to each element. At least one element comprises a second device including a means for receiving at least one control of direct access to a memory area of said second device, said control being received from said first device via said switch, and a means for transmitting said received control to a component of said second device. Said system allows said first device to perform a direct data transfer to or from a memory of said first peripheral device from or to each element.

Abstract: The subject of the invention is in particular the direct transfer of data between first and second peripherals connected via a communication bus. For this purpose, the first peripheral comprises a controller for direct access to a memory having means (425) for initiating at least one command for direct access to a region of a memory external to said first peripheral and means (400) for receiving at least one command for direct access to a region of a memory of said first peripheral, said command being received from said at least one second peripheral, and means (415) for transmitting said at least one received direct access command to a component of said first peripheral. The controller thus allows a controller of direct access to a memory of said at least one second peripheral to carry out a direct transfer of at least one data item to or from a memory of said first peripheral from or to said second peripheral.

Abstract: Methods and devices for characterizing the interconnection efficiency of a computer network based upon a static routing scheme are disclosed. The network comprises a plurality of items of equipment. After having determined (105) a plurality of types of links between said items of equipment, a theoretical number of routes per link and the associated number of links are determined (120, 125) for each type of at least one subset of said plurality of types of links according to the topology of said network. The determined theoretical number of routes per link and the associated number of links are then aggregated (135) to estimate at least one optimal routing point characterizing the theoretical efficiency of the network. Such optimal routing points can be compared with estimated actual routing points based upon the aggregation of actual numbers of routed per link and associated actual numbers of links to characterizes the network efficiency.

Abstract: A system comprising a peripheral having a timing mechanism and a node, one of which comprises a real memory space and the other a corresponding virtual memory space, is disclosed. On receiving a timing command in the real memory space, comprising references to an event and time, an entry comprising data relative to the event and time references is created in a monitoring queue of the peripheral. A current point in time is then compared, in the peripheral, to a scheduled point in time linked to an item of data relative to a time reference stored in the monitoring queue. In response, if the current point in time is after the scheduled point in time, an item of data relative to a reference linked to the item of data relative to a time reference stored in the monitoring queue is stored in the real memory space.