SERVER THAT INCLUDES A NUMEROUS OF MODULES

DE602015090554T2Active Publication Date: 2024-12-04BULL SA +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
DE602015090554
Authority / Receiving Office
DE · DE
Patent Type
Patents
Current Assignee / Owner
Priority Date
2014-08-26
Filing Date
2015-07-24
Publication Date
2024-12-04
Estimated Expiration
2035-07-24

AI Technical Summary

Technical Problem

Existing multi-module server architectures face limitations in managing a large number of interconnected modules due to the limitations of the IPMI standard, particularly in terms of sensor capacity and data logistics bus performance and flexibility, which hinders interoperability and compatibility with standard management interfaces.

Method used

A multi-module server architecture that utilizes a network of modules interconnected via an XQPI network and a private Ethernet protocol encapsulating the IPMB standard, with a BMC component capable of central management and data exchange, and a programmable gate array for clock signal distribution and error recording, ensuring software compatibility with standard management interfaces.

Benefits of technology

This solution enables efficient management of multiple modules with improved performance, reliability, and flexibility, maintaining compatibility with standard management interfaces and enhancing the resilience and scalability of the server system.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader
Need to check novelty before this filing date? Find Prior Art

Description

[0001] The present invention relates to the field of information processing and / or communication systems. More particularly, the invention proposes the production of a multi-module server and the implementation of the functions associated therewith.

[0002] Among the computer server architectures, multiprocessor network servers known as SMP, an English acronym for " Symétrie Multi Processor ", are typically implemented by a plurality of integrated circuits, commonly called sockets. For these servers, the sockets constitute the system's processors, these integrated circuits being connected by a very high-speed, very low-latency interconnection bus, allowing the creation of shared memory. The term "interconnection" will subsequently be understood to mean a physical and logical link connecting two connection interfaces.

[0003] Typically, for electrical performance reasons, each socket directly supports a portion of the SMP server's memory components. All of this memory is then made consistent using memory controllers, automatically ensuring that any data modified by one of the processors is then visible to all the others.

[0004] For physical reasons, an SMP server with a high number of sockets must be divided into several subsystems. These subsystems can, for example, be made up of daughter cards connected to a motherboard, blades connected to a backplane distributing power supplies and bus signals, or modules self-powered by the mains. This is called a multi-modular SMP server.

[0005] Regardless of the architecture of an SMP server, it requires hardware administration tools based on standards and / or norms.

[0006] Thus, the intelligent hardware management interface, called IPMI (English acronym for " Intelligent Platform Management Interface ") offers a set of interface specifications aimed at managing, that is to say supervising and controlling, the physical state of certain electronic components or hardware present in computer equipment. The IPMI standard allows, for example, monitoring in a server the regulation of temperature, voltage, power supply of microprocessors, their correct power-up, the humidity level of the components, or even the speed of the cooling fans. All of the IPMI functions can generally be activated via a basic input / output system called BIOS (an English acronym for " Basic Input / Output System ") or via management software provided by the equipment manufacturer. Thus, if necessary, for example when an alert associated with an event (e.g. overheating) of a piece of equipment is raised, the network administrator is able, via an appropriate graphical interface and from the same location (e.g. on a local or remote machine), to electrically switch off the equipment concerned, restart it, reset its parameters if necessary, or even switch on alternative equipment.

[0007] For hardware equipment (called “ hardware » in English), the supervision of the physical state of electronic components or equipment is commonly ensured by a base board management controller called BMC (English acronym for " Baseboard Management Controler "), implemented on the motherboard, or on the main board of the hardware to be supervised. For example, for a server, the functions of the BMC are carried out by the integration on the server motherboard of a system on a SOC chip (English acronym for " System On Chip "), on which a firmware (called " firmware » in English) implementing the IPMI standard.

[0008] However, the various existing implementations only address the common problem of a single-motherboard server. In the case of a server comprising a plurality of modules connected via data logistics buses to an interconnection component, for example connected according to the I 2 < C standard (English acronym for " Inter-Integrated Circuit "), the use of a BMC controller quickly becomes limited. The number of sensors increases proportionally with the number of interconnected modules, while the IPMI standard is limited to a predefined number of authorized sensors. Furthermore, increasing the number of modules requires the extension of the data logistics buses. Such an extension, for example in the case of I 2 < C, is unreliable, lacks performance and flexibility.

[0009] To circumvent the problems caused by IPMI support, one solution is to turn to proprietary administration standards. However, this solution is not very attractive if we want to continue to guarantee both interoperability and the possibility of integrating servers with existing data management systems, since these systems use standard management interfaces.

[0010] It is therefore appropriate to continue to ensure support for the IPMI standard, and more generally for standard management interfaces, for multi-module IT equipment, in order to guarantee inter-equipment compatibility.

[0011] FLYNN DAVID's US2008 / 313312 published on December 18, 2008 and KEVAN HOLDAWAY's US2009 / 276513 published on November 5, 2009 both disclose node management through BMCs included in each node but do not disclose partition formation or the use of a master clock signal.

[0012] The present invention aims to remedy the aforementioned drawbacks.

[0013] A first objective is to propose a multi-module server architecture.

[0014] A second objective is to offer a multi-module server with an expandable number of modules and allowing software compatibility with standard management interfaces.

[0015] A third objective is to provide firmware compatible with standard management interfaces, such as the IPMI standard, and allowing the management of any module constituting a multi-module server.

[0016] For this purpose, a server is proposed comprising a plurality of modules, each module comprising a communication element capable of ensuring the coherence of a shared memory between the modules; a plurality of CPU processors connected to each other and connected to the communication element; a system on a chip SOC connected to the plurality of CPU processors and to the communication element, the system on a chip SOC executing firmware; a field programmable gate array (FPGA) connected to the system on a chip SOC, to the communication element and to the plurality of CPU processors; the modules being interconnected by an interconnection between each communication element via an XQPI network; an interconnection between each system on a chip SOC via a private network in Ethernet protocol, encapsulating a communication protocol in accordance with the IPMB standard; the firmware running on each system on a SOC chip of each module realizing two software components, namely: a satellite system management controller (SMC) component capable of measuring the physical parameters of its module and managing the local functions of this module; a base board management controller (BMC) component capable of supervising all the SMCs of all the modules, centrally managing all the server functions and exchanging data with each of the BMC components via the interconnection between each system on a SOC chip.

[0017] Advantageously, in this server, the interconnection of the communication elements and the interconnection of the SOCs are carried out via an interconnection box, the interconnection box comprising a field-programmable gate array (FPGA).

[0018] Advantageously, in this server, the interconnection box is electrically powered by the modules, each module comprising at least two PSU power supplies, the PSU power supplies being sized with 2N redundancy.

[0019] Advantageously, in this server, one of the functions managed by each BMC component is a function allowing the instantiation of modules, this function being carried out in the following manner: the sending of an ID identifier request by the FPGA of each module to the interconnection box via the system on a chip SOC; the sending by the FPGA of the interconnection box of a unique ID identifier to the FPGA in response to the identifier request; the determination by the BMC component of the address of its module and the addresses of the modules to which it is interconnected, the determination of these addresses being carried out according to the ID identifier received by the FPGA.

[0020] Advantageously, in this server, the ID identifier sent by the FPGA of the interconnection box is determined according to each physical connection location of each of the modules to the interconnection module.

[0021] Advantageously, in this server, each of the modules includes a clock signal generator, capable of synchronizing the CPU processors of its module; its FPGA programmed to partition and group the modules into subsets according to common characteristics.

[0022] Advantageously, in this server, each BMC component is programmed to: identify in a set or subset of modules its belonging to a master or slave module, based on identification information of each of the modules in this set or subset; if it belongs to a master module, configure its FPGA, so that the FPGA distributes the clock signal of the master module to the slave modules of the same set or subset; if it belongs to a slave module, configure its FPGA so that the FPGA deactivates the clock signal of the slave module.

[0023] Advantageously, in this server, each CPU processor of each module comprises TSC timestamp counters, capable of synchronizing tasks comprising a plurality of light processes, the synchronization of all the TSC timestamp counters of these CPU processors in a set or a subset of module being carried out by: the sending by each BMC component of each slave module of this set or subset, of a notification to the BMC of the master module, when the slave module exits an initialization or reset phase; a notification to the BMC of the master module, when the master module exits an initialization or reset phase; the sending by the BMC of the master module, when all the master and slave modules have exited an initialization or reset phase, of a synchronization signal capable of resetting all the TSC time stamp counters of the CPU processors for all the modules of the same set or subset.

[0024] Advantageously, in this server, each FPGA of each module comprises TSC timestamp counters, capable, during any occurrence of an error in its module, of recording information relating to the error comprising at least the timestamp of the error, the synchronization of all the TSC timestamp counters of each FPGA in a set or a subset of modules being carried out by: the sending by each BMC component of each slave module of this set or subset, of a notification to the BMC of the master module, when the slave module exits an initialization or reset phase; a notification to the BMC of the master module, when the master module exits an initialization or reset phase; the sending by the BMC of the master module, when all the master and slave modules have exited an initialization or reset phase, of a synchronization signal capable of resetting all the TSC time stamp counters of the FPGAs for all the modules of the same set or subset.

[0025] Advantageously, in this server, the identification of an error in a set or subset of modules is carried out by a comparison step between the error information recorded by all the TSC time stamp counters of the FPGAs of this set or subset, the BMCs being able to exchange and share the error information in the same set or subset of modules.

[0026] Other objects and advantages of the invention will appear in the light of the description of embodiments, given below with reference to the appended drawings in which: there figure 1 is a representation of a server comprising a plurality of modules according to one embodiment; the figure 2 is a representation of the interconnection of eight modules with an interconnect box in a server according to one embodiment.

[0027] On the figure 1 an embodiment of a server is shown100 comprising a plurality of modules 1 , 2 , 3 , three in this example, such as motherboards or printed circuit boards. However, a completely different number of modules can be integrated into the server 100 , for example eight.

[0028] Advantageously, each of the modules 1 , 2 , 3 is identical and includes the following elements: one or more CPU processors 10 , 11 , 20 , 21 (English acronym for " Central Processing Unit "). In the example shown, each module 1 , 2 is bi-socket, that is, it supports two CPU processors respectively 10 , 11 , 20 , 21 . Advantageously, CPU processors 10 , 11 , 20 , 21 of the same module 1 , 2 , 3are interconnected with each other via a QPI link 101 , 201 , English acronym for " QuickPath Interconnect » ; a system on a chip SOC 12 , 22 , for example a microcontroller, running management firmware. In one embodiment, the firmware comprises two processes running in parallel, and capable of exchanging information, each respectively logically performing the function of a component: a BMC component 14 , 24 and a satellite system management controller component hereinafter referred to as the SMC component 15 , 25 , English acronym for " Satellite Management Controller " Advantageously, the functions of BMC components 14 , 24 and SMC 15 , 25 are defined in the IPMI standard. In one embodiment, the SOC 12 , 22 is able to exchange data with CPU processors 10 ,11 , 20 , 21 via a “Platform Environment Control Interface”, commonly referred to as the PECI interface 102 , 202 , English acronym for " Platform Environment Control Interface »; a “programmable gate array” component, commonly referred to as FPGA 13 , 23 (English acronym for " Field-Programmable Gate Array "), allowing in particular the transmission of signals to any component with which it is interfaced, for example clock, error and / or reset signals. Advantageously, the FPGA 13 , 23 includes a link 103 , 203 able to exchange data with CPU processors 10 , 11 , 20 , 21 and a connection 104 , 204 able to exchange data with the SOC 12 , 22 ; memories such as caches associated with CPU processors 10 , 11, 20 , 21 ; an element of communication 16 , 26 including a link 105 , 205 with each of the CPU processors 10 , 11 , 20 , 21 , capable of ensuring the consistency of a “global” memory shared between CPU processors 10 , 11 , 20 , 21 of each module 1 , 2 , 3 . In one embodiment, the communication element 16 , 26 is a BCS type component, an English acronym for “ Bull Cohérent Switch ", which is a solution proposed by the applicant and commercially available, or any later version thereof such as the BCS2 component (" Bull Cohérent Switch 2”) . The communication element 16 , 26 is also able to exchange data with the FPGA 13 , 23 via the link 103 , and with the SOC 12 , 22via bus 106 , 206 according to the I 2< C standard.

[0029] Advantageously, the modules 1 , 2 , 3 from the server 100 are interconnected by high-speed links, achieving the following interconnections: an interconnection 27 communication elements 16 , 26 modules 1 , 2 , 3 via an XQPI network, an English acronym for “ eXtended QuickPath Interconnect » , which is an interconnection solution proposed by the applicant and commercially available. Advantageously, the interconnection 27 allows communication elements 16 , 26 such as BCS2 specific to modules 1 , 2 , 3 , to exchange information, thus ensuring memory consistency across all modules 1 , 2 , 3 . The interconnection 27via the XQPI network also allows transport between each module 1 , 2 , 3 synchronization signals, such as clock signals whose synchronization is managed by FPGAs 13 , 23 ; an interconnection 28 SOCs 12 , 22 modules 1 , 2 , 3 via a private network using Ethernet protocol, encapsulating a communication protocol in accordance with the IPMB standard, the English acronym for “ Intelligent Platform Management Bus ". In the state of the art, the IPMB standard is used on communication buses, in accordance with the I 2 < C standard. The use of the IPMB standard on a LAN (English acronym for " Local Area Network "), called "IPMB on LAN", as proposed here has the advantages of improving the performance, reliability and flexibility of data transmissions between each of the SOCs 12 , 22 modules 1 , 2, 3 from the server 100.

[0030] It should be noted that each link of type 28 interconnection includes a bidirectional Ethernet link and a set of side-band signals, including SPI (Serial Peripheral Interface) signals.

[0031] In one embodiment, the interconnection 27 communication elements 16 , 26 and interconnection 28 SOCs 12 , 22 are carried out via an interconnection box 30 , illustrated on the figure 2 and described later. For example, the interconnection box 30 is a "backplane", including a network switch, commonly referred to by the Anglicism " switch » .

[0032] Advantageously, each BMC component 14 , 24 implemented on each SOC 12 , 22can be used as a data exchange interface, particularly between each BMC component 14 , 24 of each module 1 , 2 , 3 , as well as with external IT equipment management applications, such as any standard management interface. For example, an external IT equipment management application used for the server 100 , is presented in the form of a graphical interface offered to an administrator, allowing him to monitor the physical parameters of a module 1 , 2 , 3 (e.g. measured temperatures, measured voltages, detection of the presence / absence of a specific interconnection cable); communicate to the BMC component an appropriate action to be taken in the event of anomaly detection (e.g. powering on / off a module or restarting a module 1 , 2 , 3 ).

[0033] To do this, the interconnection 28 of each SOC 12 , 22 allows, from a logical point of view, the interconnection of each BMC component 14 , 24 of each module 1 , 2 , 3 . BMC components 14 , 24 of each of the modules 1 , 2 , 3 are therefore able to exchange information between them, the pooling of their functions and their information allowing the server to be presented 100 , from the perspective of an external equipment management application, such as a server 100 single-module. Thus, each BMC component 14 , 24 is programmed to manage the high-level, i.e. central, functions of the server 100 , For example : server power management 100 ; management of one or more SMC components 15 , 25, for example the monitoring of their states, the reception of measurements or any other data sent by SMC components 15 , 25 , or the transmission of control messages (piloting) to SMC components 15 , 25 ; a partitioning of clock signals based on existing partitions between different modules 1 , 2 , 3 . We subsequently designate, under the term partition in a multi-module system, a subset of modules 1 , 2 , 3 , capable of operating independently of other modules 1 , 2 , 3 of this same system. For example, a partition is a subset of modules 1 , 2 , 3 running the same operating system OS (English acronym for " Operating System "), or running the same set of operating systems OS if the modules include hypervisor layers. Advantageously, any partitioning method capable of grouping one or more sets of modules 1 , 2 , 3 sharing common characteristics, e.g. running the same operating system OS, can be used; synchronization and broadcasting, i.e. routing, of synchronization signals via the interconnect 27 through the XQPI network, based on existing partitions between different modules 1 , 2 , 3 .

[0034] Each SMC component 15 , 25 allows management of low-level, i.e. local, functions of the module 1 , 2 , 3 on which it is installed, and communicates to the BMC component 14 , 24 from the same SOC 12 , 22data it monitors, for example measurements. For example, when the server is powered on 100 , each SMC component 15 , 25 powers up its own module 1 , 2 , 3 according to a preconfigured process (e.g. sequencing, BCS2 sound programming, voltage control) and reports to the BMC component 14 , 24 from the same SOC 12 , 22 the power-on status of its module 1 , 2 , 3 . In addition, each SMC component 15 , 25 monitors the status of its own hardware, for example measures a specific number of parameters of its own module 1 , 2 , 3 (e.g. temperature, voltages) and transmits this information to the BMC component 14 , 24 from the same SOC 12 , 22 . The BMC component 14 , 24is then able to make a decision based on the parameters that the SMC component 15 , 25 communicates to him, for example restarting a module 1 , 2 , 3 especially in the event of an anomaly being detected. Thus, the hardware components of each module 1 , 2 , 3 are managed by an SMC component 15 , 25 local to the module 1 , 2 , 3 , independently of other modules 1 , 2 , 3 , while each BMC component 15 , 25 can handle a variable number of modules 1 , 2 , 3 via interconnection 28 through the IPMB on LAN. Advantageously, such an architecture is very flexible because it is independent of the number of modules 1 , 2 , 3 .

[0035] According to one embodiment, the physical topology of the Ethernet and SPI links is of the star type. The side-band signals of the clock, CATERR or TSC_SYNC type are in a mesh physical topology (called, in English, " all-to-all ”) . According to various other embodiments, other topologies are conceivable, for example Clos network topologies or multidimensional tori.

[0036] Advantageously, the interconnection box 30 is carried out according to the chosen topology. For example, in the case of a server 100 à N modules, for a topology “ all-to-all » each interconnection port of a module 1 , 2 , 3 must support N -1 high-speed links going to each of its neighbors. If in an eight-module configuration, where N =8, each module 1 , 2 , 3includes seven eight-way unidirectional links, the interconnect box 30 is then made to support 8*7*8, or 448 unidirectional high-speed lanes.

[0037] In addition, in order to improve the resilience of the interconnect box 30 In case of failure, it is possible to realize the switch of this box via a plurality of switches. For example, for a server 100 with eight modules, in order to interconnect the SOCs 12 , 22 modules 1 , 2 , 3 , the switch of the interconnection box 30 is an eight-port Ethernet switch. Such a switch can then be implemented via two five-port switches as follows: if each of the switches includes four data downlink ports (so-called " downlink ") and a data uplink port (called "port uplink "), we then connect the ports " uplink » via a printed circuit board trace. Thus, SOCs 12 , 22 modules 1 , 2 , 3 can only lose all of their means of dialogue in the event of failure of all the switches, this risk decreasing proportionally given the number of switches used to create the interconnection box switch 30.

[0038] For the same reasons of resilience, all the hardware components that provide the physical links between the modules are passive components. Advantageously, these passive components are chosen so as to have a "mean time between failures" MTBF (an English acronym for " mean time between failures » several orders of magnitude higher than the MTBF of the modules 1 , 2 , 3.Thus, the number of failures created by these components is negligible when trying to assess server availability. 100.

[0039] Always with a view to ensuring good resilience against power supply failures, each module 1 , 2 , 3 includes at least two PSU power supplies (acronym for " Power Supply Unit "), which can, for example, be presented in the form of AC / DC converters (English acronym for " Alternating Current / Direct Current "), delivering an input power of each module 1 , 2 , 3 , for example under 12V. Each of the PSU power supplies can also be sized with 2N redundancy, i.e. include a doubling of their technical components (e.g. electronic components), and can be connected to two independent AC alternating current networks, ensuring the operation of the server100 if one of the networks is faulty. The interconnection box 30 is electrically powered by the modules 1 , 2 , 3 , each sending across the interconnect's Ethernet links 28 a current under a predefined voltage, for example 12V. These currents are then summed by an appropriate electronic device constituting the interconnection box 30 , for example, a diode switch. In the state of the art, the PoE standard, an English acronym for " Power-over-Ethernet » , describes the power supply of devices interfaced via an Ethernet link with a switch, thanks to the switch power supply. A power failure of the switch therefore prevents any power supply to the devices. In contrast to this standard, the embodiment described above makes it possible to reduce the impact of a power failure on the server 100, these being then limited to the scope of a possible breakdown of a module 1 , 2 , 3 , such a probability of failure being reduced thanks to the redundancy of the power supplies. Thus, all Ethernet links made via the interconnection 27 modules 1 , 2 , 3 collaborate to provide redundant power to the interconnect box switch 30 , only this switch including active components. The switch of the interconnection box 30 is therefore capable of functioning despite a possible failure of one or more modules 1 , 2 , 3. Advantageously, the power dissipated by this switch is very low (of the order of 2 to 3 Watts) and the number of components present in this switch makes it possible to provide an MTBF of the order of several million hours. In addition, care is taken to configure all the BMC components 14 , 24 SOCs 12 , 22 to supervise the operation of this switch and manage any malfunction. Advantageously, the embodiments described above allow any module 1 , 2 , 3 to withstand a failure of the interconnection box switch 30 , despite its central position. If a failure is detected on the switch, hot-swapping this module does not interrupt server operation. 100 multi-modular.

[0040] There figure 2 illustrates the interconnection of eight modules 1 , 2 , 3, 4 , 5 , 6 , 7 , 8 with an interconnection box 30 , according to one embodiment, and in accordance with the characteristics previously described. In this figure, each module 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 includes respectively the previously mentioned elements, in particular: an FPGA 13 , 23 , 33 , 43 , 53 , 63 , 73 , 83 able to manage a clock signal 130 , 230 , 330 , 430 , 530 , 630 , 730 , 830 specific. The management of these signals is detailed later; a BMC component 14 , 24 , 34 , 44 , 54 , 64 , 74 , 84able to handle high-level functions such as instantiation, power-up, reboot and module instantiation 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 .

[0041] Communication elements 16, 26, CPU processors 10, 11, 20, 21, the SOCs 12, 22 and interconnections 27, 28 are not shown in this figure to simplify its reading, but are nevertheless present. In particular, each module 1, 2, 3, 4, 5, 6, 7, 8 is respectively interconnected by its communication element 16, 26 and its SOC 12, 22 to other modules 1, 2, 3, 4, 5, 6, 7, 8, respectively via the interconnections 27, 28 and through the interconnection box 30.

[0042] In one embodiment, the connection of each module 1, 2, 3, 4, 5, 6, 7, 8 to the interconnection box 30is carried out at the latter level respectively via a synchronous SPI data bus 131, 231, 331, 431, 531, 631, 731, 831, English acronym for " Serial Peripheral Interface » . Data exchanges between each module 1, 2, 3, 4, 5, 6, 7, 8 and the interconnection box 30 via the SPI synchronous data buses 131, 231, 331, 431, 531, 631, 731, 831 are represented in this figure by double arrows between these elements. In addition, each of the SPI synchronous data buses 131, 231, 331, 431, 531, 631, 731, 831 is respectively interfaced with a register 132, 232, 332, 432, 532, 632, 732, 832 of an additional device of the interconnection box 30, here an FPGA. Advantageously, the FPGA of the interconnection box 30 will allow to allocate to each SOC 12, 22 of each module 1, 2, 3, 4, 5, 6, 7, 8 an identifier during an initialization phase. Information exchanges between SPI synchronous data buses 131, 231, 331, 431, 531, 631, 731, 831 and the registers 132, 232, 332, 432, 532, 632, 732, 832are symbolized here by double arrows between these elements. A module initialization phase 1, 2, 3, 4, 5, 6, 7, 8 may, for example, occur during a power-up, a restart of the server 100 or even when initializing a SOC 12, 22 by a BMC component 14, 24, 34, 44, 54, 64, 74, 84.

[0043] As explained previously, an interconnection 28 of each SOC 12, 22 of each module 1, 2, 3, 4, 5, 6, 7, 8 is carried out via the communication box 30 through a private network in Ethernet protocol, encapsulating a communication protocol in accordance with the IPMB standard. Each SOC 12, 22 must therefore be able to start its IP (Internet Protocol) connection during an initialization phase via an address different from that of its neighbors. To do this, according to various embodiments, each FPGA 13, 23, 33, 43, 53, 63, 73, 83 sends via SOC 12, 22with which it is interfaced, via the interconnection 28, an identifier request to the interconnection box 30. Each request is received by the communication box 30, at the SPI synchronous data bus level 131, 231, 331, 431, 531, 631, 731, 831, the latter transmitting each request to the registry 132, 232, 332, 432, 532, 632, 732, 832 of the FPGA with which it is interfaced. Advantageously, each request allows obtaining a unique ID identifier known only to the register 132, 232, 332, 432, 532, 632, 732, 832 of the FPGA. This ID identifier is delivered by means of the SPI bus signals which are physically independent of the Ethernet signals (of course, the side-band signals are not an integral part of the XQPI interface).

[0044] In one embodiment, each ID identifier relates to a port number, coded for example on three bits, this ID identifier being specific to each physical connection location to the interconnection module. 30.Advantageously, this ID identifier is unique and therefore allows a module to be identified subsequently 1, 2, 3, 4, 5, 6, 7, 8 specific. The communication box 30, then communicates, through its FPGA, an ID identifier in response to each request received, this response being transmitted via the SPI synchronous data bus 131, 231, 331, 431, 531, 631, 731, 831 then by the SOC 12, 22 to the FPGA 13, 23, 33, 43, 53, 63, 73, 83 of the module 1, 2, 3, 4, 5, 6, 7, 8 having issued the request. The ID identifier received by the FPGA 13, 23, 33, 43, 53, 63, 73, 83, is then read by the BMC component 14, 24, 34, 44, 54, 64, 74, 84 of the SOC 12, 22 with which it is interfaced in order to determine an appropriate instantiation, for example the assignment of an IP address to the module 1, 2, 3, 4, 5, 6, 7, 8. Reading the ID identifier by each BMC component 14, 24, 34, 44, 54, 64, 74, 84 in each FPGA 13, 23, 33, 43, 53, 63, 73, 83 is symbolized in this figure by a one-sided arrow between these elements. In one embodiment, each BMC component 14, 24, 34, 44, 54, 64, 74, 84applies an IP address calculation algorithm to this ID identifier in order to dynamically determine and instantiate the IP address of its module 1, 2, 3, 4, 5, 6, 7, 8. Furthermore, each FPGA 13, 23, 33, 43, 53, 63, 73, 83 is capable of reading via the SPI synchronous data bus 131, 231, 331, 431, 531, 631, 731, 831 the number of modules 1, 2, 3, 4, 5, 6, 7, 8 connected to the interconnect box. Thus, the BMC component 14, 24, 34, 44, 54, 64, 74, 84 of each module 1, 2, 3, 4, 5, 6, 7, 8 is also able to deduce the IP address of the modules 1, 2, 3, 4, 5, 6, 7, 8 neighbors via the application of the same IP address calculation algorithm. All determined IP addresses can then, for example, be stored by the 13, 23, 33, 43, 53, 63, 73, 83.

[0045] In another embodiment, the IP addresses of each module 1, 2, 3, 4, 5, 6, 7, 8are IPv4 type addresses, consisting of 32 bits. It is further assumed that each IP address has twenty-nine common most significant bits, such a configuration being applicable in a private Ethernet network such as that of the interconnection 28. After retrieving each ID identifier by each of the FPGAs 13, 23, 33, 43, 53, 63, 73, 83, each BMC component 14, 24, 34, 44, 54, 64, 74, 84 then completes this IP address by completing the three remaining low-weight bits with the three bits of the ID identifier that it will have read at the level of its FPGA 13, 23, 33, 43, 53, 63, 73, 83. So each BMC component 14, 24, 34, 44, 54, 64, 74, 84 can be seen as an interface for dynamic configuration or reconfiguration (i.e. self-configuration) of modules 1, 2, 3, 4, 5, 6, 7, 8 allowing in particular their instantiation, that is to say the allocation of an IP address or more generally an identifier to each of these modules 1 , 2 , 3 , 4 , 5 , 6 , 7 ,8.

[0046] Each module 1, 2, 3, 4, 5, 6, 7, 8 further comprises a clock generator, capable of generating a clock signal 130 , 230 , 330 , 430, 530, 630, 730, 830 which is specific to it. Such a clock is, for example, produced in the form of a square signal whose frequency is of the order of MHz, for example 14.7 MHz, 25 MHz or 100 MHz. Advantageously, a clock signal 130 , 230 , 330 , 430 , 530 , 630 , 730, 830 in a module 1, 2, 3, 4, 5, 6, 7, 8 allows synchronization of all CPU processors 10, 11, 20, 21 of this module 1, 2, 3, 4, 5, 6, 7, 8. The distribution of each clock signal 130 , 230 , 330, 430, 530, 630, 730, 830 to CPU processors 10, 11, 20, 21 of each module 1, 2, 3, 4, 5, 6, 7, 8, is represented here by the one-way arrows to the left of the elements 130 , 230 , 330 , 430 , 530, 630, 730, 830.

[0047] According to various embodiments, each clock signal 130 , 230, 330, 430, 530, 630, 730, 830 can be transmitted through the interconnection 27 via the XQPI network. Advantageously, each clock signal 130 , 230 , 330 , 430 , 530 , 630 , 730 , 830 can be filtered at the input of each module 1, 2, 3, 4, 5, 6, 7, 8 by an electronic circuit of the phase-locked loop type PLL, English acronym for " Phase-Locked Loop ", capable of eliminating any phase noise introduced by the physical links of the interconnection 27.

[0048] Clock signals 130 , 230 , 330 , 430 , 530 , 630 , 730 , 830 which can be inter-module signals, i.e. passing between different modules 1, 2, 3, 4, 5, 6, 7, 8 synchronization of these signals is also necessary.

[0049] To do this, according to various embodiments, each FPGA 13, 23, 33, 43, 53, 63, 73, 83 of each module 1, 2, 3, 4, 5, 6, 7, 8 is able, during a partitioning step, to group all the modules 1, 2, 3, 4, 5, 6, 7, 8 based on common characteristics, for example based on the same OS operating system supported by the modules 1, 2, 3, 4, 5, 6, 7, 8, thus forming partitions. Each module 1, 2, 3, 4, 5, 6, 7, 8 therefore belongs to a set or partition (i.e. a subset) of modules 1, 2, 3, 4, 5, 6, 7, 8, server constituent 100. For example, on the figure 2 , two partitions 40, 50 are carried out: the first partition 40 is formed from modules 1, 2, 3, 4, 5, 6 and the second partition 50 is formed from modules 7, 8.

[0050] Each BMC component 14, 24, 34, 44, 54, 64, 74, 84 then compares in the set or subset (i.e. partition) to which it belongs, the number of its own module 1, 2, 3, 4, 5, 6, 7, 8with the module numbers 1, 2, 3, 4, 5, 6, 7, 8 of the same set or subset. As previously stated, the number of a module 1, 2, 3, 4, 5, 6, 7, 8 and its modules 1, 2, 3, 4, 5, 6, 7, 8 neighbors is identified, during a dynamic instantiation step carried out by its BMC component 14, 24, 34, 44, 54, 64, 74, 84.

[0051] Based on the result of these comparisons, each BMC 14, 24, 34, 44, 54, 64, 74, 84 is able to identify whether the module 1, 2, 3, 4, 5, 6, 7, 8 to which it belongs is a master or slave module in the set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8. In one embodiment, the module 1, 2, 3, 4, 5, 6, 7, 8 lowest number / ID identifier is identified by its BMC 14, 24, 34, 44, 54, 64, 74, 84as the master module, while the remaining modules are identified as slave modules. However, any other type of identification can be performed, for example the master module can be identified as the module with the highest ID number, or identified based on its address. More generally, master and slave modules are identified within the same set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8, based on module credentials 1, 2, 3, 4, 5, 6, 7, 8, obtained during their dynamic instantiation.

[0052] For the sake of simplification, we will subsequently consider the example in which the module 1, 2, 3, 4, 5, 6, 7, 8 lowest numbered in an assembly or subassembly is identified by each BMC component 14, 24, 34, 44, 54, 64, 74, 84 as the master module, while the remaining modules are identified as slave modules.

[0053] According to various embodiments, if the BMC component 14, 24, 34, 44, 54, 64, 74, 84identifies its module 1, 2, 3, 4, 5, 6, 7, 8 as being a module master, the BMC component 14, 24, 34, 44, 54, 64, 74, 84 configures its FPGA 13, 23, 33, 43, 53, 63, 73, 83 so that it distributes its clock signal 130, 230, 330, 430, 530, 630, 730, 830 to other modules 1, 2, 3, 4, 5, 6, 7, 8 of the same set or subset; slave, the BMC component 14, 24, 34, 44, 54, 64, 74, 84 configures its FPGA 13, 23, 33, 43, 53, 63, 73, 83 so that it deactivates the clock signal 130, 230, 330, 430, 530, 630, 730, 830 local to this module 1, 2, 3, 4, 5, 6, 7, 8.

[0054] For example, in the same set of eight modules 1, 2, 3, 4, 5, 6, 7, 8 non-partitioned, the module 1 is identified as the master module and the modules 2, 3, 4, 5, 6, 7, 8 as slave modules. The module 1 then distributes its clock signal 130 to the modules 2, 3, 4, 5, 6, 7, 8, the latter having their clock signal 230, 330, 430, 530, 630, 730, 830 disabled by their FPGA. The clock signal distribution 130 and local deactivation of clock signals 230, 330, 430, 530, 630, 730, 830are here carried out respectively by the FPGA 13 and FPGAs 23, 33, 43, 53, 63, 73, 83 these being respectively configured by the BMC component 14 and BMC components 24, 34, 44, 54, 64, 74, 84. The clock is transmitted by a sideband signal, and is therefore part of the interconnection 28 (i.e. the inter FPGA interface).

[0055] In another example illustrated on the figure 2 : the module 1 is identified as the master module in the first partition 40, the modules 2, 3, 4, 5, 6 being identified as slave modules in this partition. The module 1 then distributes through its FPGA 13 (one-way arrows coming from this element), its clock signal 130 to the modules 2, 3, 4, 5, 6, the latter having deactivated locally via their FPGA 23, 33, 43, 53, 63 their clock signal 130, 230, 330, 430, 530, 630(one-way arrows in the direction from FPGA to clock signal). Advantageously, the FPGA configuration 13 and FPGAs 23, 33, 43, 53, 63 is carried out respectively by the BMC component 14 and BMC components 24, 34, 44, 54, 64 (one-way arrows in the direction of clock signal to FPGA); the module 7 is identified as the master module in the second partition 50, the module 8 being identified as a slave module in this partition. The module 7 then distributes through its FPGA 73 (one-way arrow coming from this element), its clock signal 730 to the module 8, the latter having locally deactivated its clock signal 830 via its FPGA 63 (one-way arrow in the direction of FPGA to clock signal). Advantageously, the FPGA configuration 73 and the FPGA 83,is carried out respectively by the BMC component 74 and the BMC component 84 (one-way arrow in the direction of clock signal to FPGA).

[0056] Each CPU processor 10, 11, 20, 21 of each module 1, 2, 3, 4, 5, 6, 7, 8 is therefore driven by a clock signal 130, 230, 330, 430, 530, 630, 730, 830 common to the score or set of modules 1, 2, 3, 4, 5, 6, 7, 8 to which it belongs. Advantageously, each CPU processor 10, 11, 20, 21 comprises a plurality of registers, each register relating to a lightweight process (“ thread » in English) executed by the processor 10, 11, 20, 21. These registers are commonly referred to as TSC (Time Stamp Counters) Time Stamp Counter "), and are used to synchronize tasks " multi-threads ", that is, comprising a plurality of lightweight processes. The TSC timestamp counters are initialized for each CPU processor 10, 11, 20, 21 after an initialization / reset step (“ reset » in English) which may, for example, occur when powering up / restarting a module 1, 2, 3, 4, 5, 6, 7, 8. For proper execution of tasks " multi-threads ", the TSC timestamp counters must therefore be perfectly synchronized. Advantageously, the synchronization of the TSC timestamp counters is carried out via clock signals, 130, 230, 330, 430, 530, 630, 730, 830 at the input of CPU processors 10, 11, 20, 21. Such synchronization is complex to ensure, particularly for modules 1, 2, 3, 4, 5, 6, 7, 8 interconnected each comprising a plurality of processors 10, 11, 20, 21, because each CPU processor 10, 11, 20, 21 may initially and locally have its own clock signal 130, 230, 330, 430, 530, 630, 730, 830. TSC timestamp counters between each module 1, 2, 3, 4, 5, 6, 7, 8 are therefore potentially inconsistent from a temporal point of view and may also drift.

[0057] Thus, according to various embodiments, after a possible partitioning step of the modules 1, 2, 3, 4, 5, 6, 7, 8, synchronization of TSC timestamp counters of CPU processors 10, 11, 20, 21 is carried out as follows: each BMC component 14, 24, 34, 44, 54, 64, 74, 84 of each master module of a set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8 performs FPGA configuration 13, 23, 33, 43, 53, 63, 73, 83 of the master module, so that the routing of any synchronization signal is restricted to the slave modules of the same set or subset; each BMC component 14, 24, 34, 44, 54, 64, 74, 84 of each slave module sends, via the interconnection 28 through the IPMB network on LAN, a notification message to the BMC component 14, 24, 34, 44, 54, 64, 74, 84 of the master module when the slave module exits an initialization or reset phase (corresponding to an initialization of the TSC timestamp counters). The BMC component 14, 24, 34, 44, 54, 64, 74, 84The master module can also be informed via a notification of the initialization or reinitialization of its own module. 1, 2, 3, 4, 5, 6, 7, 8 ; the BMC component 14, 24, 34, 44, 54, 64, 74, 84 of the master module waits for all modules 1, 2, 3, 4, 5, 6, 7, 8, including itself, are out of the initialization or reset phase; when notifications from all modules 1, 2, 3, 4, 5, 6, 7, 8 were received, the BMC 14, 24, 34, 44, 54, 64, 74, 84 of the master module triggers a reset of all TSC timestamp counters of the CPU processors 10, 11, 20, 21 for all modules 1, 2, 3, 4, 5, 6, 7, 8 of the same assembly or subassembly to which it belongs, including for its module. Advantageously, this reset is carried out by sending a synchronization signal on a physical connection pin (" pin » in English) that each CPU processor has 10, 11, 20, 21 of each module 1, 2, 3, 4, 5, 6, 7, 8.Sending such a synchronization signal therefore causes a synchronous restart of all the TSC timestamp counters of the CPU processors 10, 11, 20, 21 and therefore their temporal consistency, because each CPU processor 10, 11, 20, 21 then starts TSC counters controlled by the same clock in the set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8.

[0058] Furthermore, as stated in the description of the figure 1 , each FPGA 13, 23, 33, 43, 53, 63, 73, 83 of each module 1, 2, 3, 4, 5, 6, 7, 8 includes a data link with each of the CPU processors 10, 11, 20, 21 as well as with a communication element 16, 26 such as a BCS2. When a first uncorrected, fatal or catastrophic error occurs at the level of a CPU processor 10, 11, 20, 21 or a communication element 16, 26 of a module 1, 2, 3, 4, 5, 6, 7, 8, This can propagate at very high speed, generally in a few microseconds, via the interconnection 27through the XQPI network, to other modules 1, 2, 3, 4, 5, 6, 7, 8 of the same set or subset, generating errors in their CPU processors 10, 11, 20, 21 and their communication elements 16, 26. In a multi-modular context, it is therefore necessary to be able to find the first error, in order to diagnose the fault. More generally, it is also necessary to be able to identify (e.g., locate, date) precisely any occurrence of an error in a module. 1, 2, 3, 4, 5, 6, 7, 8.

[0059] Thus, according to various embodiments, we also use for each FPGA 13, 23, 33, 43, 53, 63, 73, 83 a TSC timestamp counter, implemented by a register of configurable size, for example forty bits, and synchronized via the clock signal 130, 230, 330, 430, 530, 630, 730, 830 of the master module of the set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8 to which the FPGA belongs 13, 23, 33, 43, 53, 63, 73, 83, for example at a frequency of 25MHz. Each of the FPGAs 13, 23, 33, 43, 53, 63, 73, 83therefore has a TSC timestamp counter perfectly synchronized with those of the FPGAs 13, 23, 33, 43, 53, 63, 73, 83 of the same set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8. Advantageously, thanks to this synchronous time reference, each FPGA 13, 23, 33, 43, 53, 63, 73, 83 is able to date (“ timestamp ”) and record a report regarding any error or event that occurs in its module 1, 2, 3, 4, 5, 6, 7, 8. It will therefore be possible at any time to reconstruct the chronology of several errors or events without ambiguity about their respective order. To do this, following the timestamping by the FPGAs 13, 23, 33, 43, 53, 63, 73, 83 events, for example errors propagated in each module 1, 2, 3, 4, 5, 6, 7, 8, all BMC components 14, 24, 34, 44, 54, 64, 74, 84 are able to exchange and share information relating to these events, making it possible in particular to find the source of these errors, i.e. the first error. The precise location and diagnosis of a module 1, 2, 3, 4, 5, 6, 7, 8defaulting are therefore facilitated.

[0060] According to various embodiments, after a possible partitioning step of the modules 1, 2, 3, 4, 5, 6, 7, 8, synchronization of FPGA TSC timestamp counters 13, 23, 33, 43, 53, 63, 73, 83, and the detection of a first error is carried out as follows: each BMC component 14, 24, 34, 44, 54, 64, 74, 84 of each master module of a set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8 performs FPGA configuration 13, 23, 33, 43, 53, 63, 73, 83 of the master module, so that the routing of any synchronization signal is restricted to the slave modules of the same set or subset; each BMC component 14, 24, 34, 44, 54, 64, 74, 84 of each slave module sends, via the interconnection 28 through the IPMB network on LAN, a notification message to the BMC component 14, 24, 34, 44, 54, 64, 74, 84of the master module when the slave module exits an initialization or reset phase (corresponding to an initialization of the TSC timestamp counters). The BMC component 14, 24, 34, 44, 54, 64, 74, 84 The master module can also be informed via a notification of the initialization or reinitialization of its own module. 1, 2, 3, 4, 5, 6, 7, 8 ; the BMC component 14, 24, 34, 44, 54, 64, 74, 84 of the master module waits for all modules 1, 2, 3, 4, 5, 6, 7, 8, including itself, are out of the initialization or reset phase; when notifications from all modules 1, 2, 3, 4, 5, 6, 7, 8 were received, the BMC 14, 24, 34, 44, 54, 64, 74, 84 of the master module triggers a reset of all the TSC timestamp counters of the FPGAs 13, 23, 33, 43, 53, 63, 73, 83 of all modules 1, 2, 3, 4, 5, 6, 7, 8 of the same set or subset to which it belongs, including for itself. The timestamp counters of the different FPGAs 13, 23, 33, 43, 53, 63, 73, 83are then perfectly synchronized; at each occurrence of an error (or an event) occurring in the module 1, 2, 3, 4, 5, 6, 7, 8 , for example at the level of a CPU processor 10, 11, 20, 21 or a communication element 16, 26 (eg: a BCS2), the FPGA 13, 23, 33, 43, 53, 63, 73, 83 of this module 1, 2, 3, 4, 5, 6, 7, 8 stores error information via its TSC timestamp counter. Error information includes, for example, its type (e.g., uncorrected, fatal), its source (e.g., module number, type of component involved: CPU or BCS), and its timestamp; each BMC component 14, 24, 34, 44, 54, 64, 74, 84 of each module 1, 2, 3, 4, 5, 6, 7, 8, then accesses the information stored by the TSC timestamp counter in the FPGA 1, 2, 3, 4, 5, 6, 7, 8 and appropriately makes this information accessible to any server administrator 100 or external tool allowing server management 100. For example, each BMC component 14, 24, 34, 44, 54, 64, 74, 84communicates the information read from the FPGA 1, 2, 3, 4, 5, 6, 7, 8 via a log file accessible from an external equipment management application; a comparison step between the information, for example between the timestamp, the source, the type of each error, communicated by each BMC component 14, 24, 34, 44, 54, 64, 74, 84 then allows to identify a first error among a set of errors occurring in different modules 1, 2, 3, 4, 5, 6, 7, 8, as well as the chronology of subsequent errors. Such a comparison step can, for example, be automated or performed directly by the server administrator 100. Advantageously, the identification, for example the determination of the chronology and the location, of the first error subsequently makes it possible to decide on one or more actions to be carried out on the server. 100, for example restart a module 1, 2, 3, 4, 5, 6, 7, 8 specific, or reconfigure a partition if a module 1, 2, 3, 4, 5, 6, 7, 8in this one is faulty.

[0061] Advantageously, the embodiments described above allow software compatibility with standard management interfaces, for example interfaces in accordance with the IPMI standard, regardless of the number of modules. 1, 2, 3, 4, 5, 6, 7, 8 server components 100. This is made possible in particular thanks to: to BMC components 14, 24, 34, 44, 54, 64, 74, 84, each one being usable as a data communication interface, allowing in particular the synchronization of clocks, TSC time stamp counters, as well as the management of high-level functions, i.e. the central functions, of the different modules 1, 2, 3, 4, 5, 6, 7, 8 ; to SMC components 15, 25 in charge of managing low-level functions, i.e. local to their own modules 1, 2, 3, 4, 5, 6, 7, 8, for example the measurement of their physical parameters and their powering up. This provides great flexibility in terms of hardware and software architecture, facilitating any possible future need for modification, adaptation or evolution of the server. 100.

[0062] Advantageously, the embodiments described above allow the management of a set or a subset of modules. 1, 2, 3, 4, 5, 6, 7, 8 thanks to a partitioning of these latter, as well as of their clock signals 230, 330, 430, 530, 630, 730, 830 respective. The management of a set or subset of modules 1, 2, 3, 4, 5, 6, 7, 8 is particularly advantageous, because it allows the support of different OS depending on the modules 1, 2, 3, 4, 5, 6, 7, 8, the identification, scheduling and precise localization of any possible errors for diagnostic purposes, as well as possible dynamic reconfiguration of partitions, particularly in the event of module failure 1, 2, 3, 4, 5, 6, 7, 8.

[0063] Another advantage of the proposed implementations lies in the hardware implementation of the server. 100, and in particular, the way in which the interconnection box 30 and the modules 1, 2, 3, 4, 5, 6, 7, 8 are interconnected and electrically powered, allows hot maintenance of each element of the server 100 as well as an improvement in its resilience.

[0064] Further, the previously described embodiments for a server 100 are also transposable to any IT equipment or infrastructure comprising a plurality of modules 1, 2, 3, for example to a supercomputer.

Claims

1. A server (100) comprising a plurality of modules (1-8), each module (1-8) comprising: - a communication element (16, 26) capable of ensuring the consistency of a memory shared between modules (1-8); - a plurality of CPU processors (10, 11, 20, 21) connected to each other and to the communication element (16, 26); - a system-on-a-chip SOC (12, 22) connected to the plurality of CPU processors (10, 11, 20, 21) and to the communication element (16, 26), the system-on-a-chip SOC (12, 22) running firmware; - a field-programmable gate array (FPGA) (13, 23, 33, 43, 53, 63, 73, 83) connects to the system-on-a-chip SOC (12, 22), the communication element (16, 26) and the plurality of CPU processors (10, 11, 20, 21); the modules (1-8) being interconnected by: - an interconnection (27) between each communication element (16, 26) via an XQPI (extended QuickPath Interconnect) network; - an interconnection (28) between each system-on-a-chip SOC (12, 22) via a private Ethernet protocol network, encapsulating a communication protocol in accordance with the IPMB standard; the firmware running on each system-on-a-chip SOC (12, 22) of each module (1-8) implementing two software components, namely: - a satellite management controller (SMC) component (15, 25) of the system, capable of measuring the physical parameters of its module (1-8) and managing the local functions of this module (1-8); - a baseboard management controller (BMC) component (14, 24, 34, 44, 54, 64, 74, 84) capable of supervising all the SMCs (15, 25) of all the modules (1-8), in particular by receiving measurements or any other data sent back by said SMCs and by transmitting control messages to said SMCs, centrally managing all the functions of the server (100), and exchanging data with each of the BMC components (14, 24, 34, 44, 54, 64, 74, 84) via the interconnection (28) between each system-on-a-chip SOC (12, 22), each of the modules (1-8) further comprising a clock signal generator (130, 230, 330, 430, 530, 630, 730, 830), adapted to synchronize the CPU processors (10, 11, 20, 21) of its module (1-8); and its FPGA (13, 23, 33, 43, 53, 63, 73, 83) being programmed to partition and group into subsets the modules (1-8) according to a same operating system supported by said modules; and said interconnection (27) via an XQPI network being configured to route said clock signal between different modules (1-8) depending on said subsets.

2. The server (100) according to claim 1, wherein the interconnection (27) of the communication elements (16, 26) and the interconnection (28) of the SOCs (12, 22) are made via an interconnection box (30), the interconnection box (30) comprising a field programmable gate array (FPGA).

3. The server (100) according to claim 2, wherein the interconnection box (30) is supplied by the modules (1-8), each module (1-8) comprising at least two power supply units PSUs sized with 2N redundancy.

4. The server (100) according to claim 2 or 3, wherein one of the functions managed by each BMC component (14, 24, 34, 44, 54, 64, 74, 84) is a function enabling the instantiation of modules (1-8), for which purpose: - the FPGA (13, 23, 33, 43, 53, 63, 73, 83) of each module (1-8) being configured to send an ID request for each module (1-8) to the interconnection box (30) via the system-on-a-chip SOC (12, 22); - the FPGA of the interconnect box (30) being configured to send a unique identifier ID to the FPGA (13, 23, 33, 43, 53, 63, 73, 83) in response to the ID request; - the BMC component (14, 24, 34, 44, 54, 64, 74, 84) being configured to determine the address of its module (1-8) and the addresses of the modules (1-8) to which it is interconnected, these addresses being determined as a function of the identifier ID received by the FPGA (13, 23, 33, 43, 53, 63, 73, 83).

5. The server (100) according to claim 4, wherein the identifier ID sent by the FPGA of the interconnection box (30) is determined as a function of each physical connection location of each of the modules (1-8) to the interconnection module (30).

6. The server (100) according to any of the preceding claims, wherein each BMC component (14, 24, 34, 44, 54, 64, 74, 84) is programmed to: - identify, in a set or subset of modules (1-8), whether it belongs to a master or slave module, based on identification information for each of the modules (1-8) in that set or subset; - if it belongs to a master module, configure its FPGA (13, 23, 33, 43, 53, 63, 73, 83) so that the FPGA (13, 23, 33, 43, 53, 63, 73, 83) distributes the clock signal (130, 230, 330, 430, 530, 630, 730, 830) from the master module to the slave modules in the same set or subset; - if it belongs to a slave module, configure its FPGA (13, 23, 33, 43, 53, 63, 73, 83) so that the FPGA (13, 23, 33, 43, 53, 63, 73, 83) disables the clock signal (130, 230, 330, 430, 530, 630, 730, 830) of the slave module.

7. The server (100) according to claim 6, wherein each CPU processor (10, 11, 20, 21) of each module (1-8) comprises timestamp counters TSC, suitable for synchronizing tasks comprising a plurality of light processes, for synchronizing all the timestamp counters TSC of these CPU processors (10, 11, 20, 21) in a module (1-8) set or subset: - each BMC component (14, 24, 34, 44, 54, 64, 74, 84) of each slave module of that set or subset being configured to send a notification to the BMC (14, 24, 34, 44, 54, 64, 74, 84) of the master module, when the slave module exits an initialization or reset phase; - The BMC (14, 24, 34, 44, 54, 64, 74, 84) of the master module is notified when the master module exits an initialization or reset phase; - The BMC (14, 24, 34, 44, 54, 64, 74, 84) of the master module being configured to send, when all the master and slave modules have exited an initialization or reset phase, a synchronization signal capable of resetting all the timestamp counters TSC of the CPU processors (10, 11, 20, 21) for all the modules (1-8) in the same set or subset.

8. The server (100) according to claim 6 or 7, wherein each FPGA (13, 23, 33, 43, 53, 63, 73, 83) of each module (1-8) comprises timestamp counters TSC, capable on any occurrence of an error in its module (1-8) of recording information relating to the error comprising at least the timestamp of the error, for the synchronization of the set of timestamp counters TSC of each FPGA (13, 23, 33, 43, 53, 63, 73, 83) in a module (1-8) set or subset: - each BMC component (14, 24, 34, 44, 54, 64, 74, 84) of each slave module of that set or subset being configured to send a notification to the BMC (14, 24, 34, 44, 54, 64, 74, 84) of the master module, when the slave module exits an initialization or reset phase; - The BMC (14, 24, 34, 44, 54, 64, 74, 84) of the master module is notified when the master module exits an initialization or reset phase; - The BMC (14, 24, 34, 44, 54, 64, 74, 84) of the master module being configured to send, when all the master and slave modules have exited an initialization or reset phase, a synchronization signal capable of resetting all the timestamp counters TSC of the FPGAs (13, 23, 33, 43, 53, 63, 73, 83) for all the modules (1-8) in the same set or subset.

9. The server (100) according to claim 8, comprising, for identifying an error in a set or subset of modules (1-8), means configured to perform a comparison between the error information recorded by the set of timestamp counters TSC of the FPGAs (13, 23, 33, 43, 53, 63, 73, 83) of that set or sub-set, the BMCs (14, 24, 34, 44, 54, 64, 74, 84) being able to exchange and share error information within the same set or sub-set of modules (1-8).