Operating method of automatic extracting idle computing GPU resource with clustering for AI data processing and system thereof
The cluster computing system addresses the inefficiency in utilizing idle GPU resources by automatically detecting and managing them for AI tasks, forming a high-performance computing infrastructure that meets AI demands efficiently and cost-effectively.
Patent Information
- Application Number
- KR1020250174044
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2045-11-17
AI Technical Summary
Conventional systems fail to identify and utilize high-performance GPU resources efficiently for artificial intelligence computations due to the lack of real-time collection and management of GPU specifications, processing capacity, and available time slots, leading to wasted resources.
A cluster computing system that automatically detects and manages idle GPU resources by regularly collecting operational information in real time, forming a distributed AI cluster with required specifications and capacity, and providing it at the needed time for large-scale AI computations.
Enables rapid and cost-effective assembly of a high-performance computing infrastructure for AI tasks by continuously identifying and integrating idle GPU resources, ensuring efficient and timely processing capabilities.
Smart Images

Figure 112025128555412-PAT00007_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a cluster computing system configured for the automatic detection of idle computing GPU resources for artificial intelligence data processing and a method of operating the same. More specifically, the invention relates to a cluster computing system configured for the automatic detection of idle computing GPU resources for artificial intelligence data processing and a method of operating the same, which regularly collects the specifications and operational information of each high-performance computer connected to a network in real time through an agent, searches for GPU resources among unused idle computing resources, connects them into a cluster, and operates them as a single computer, thereby establishing and operating an infrastructure in real time that enables rapid and efficient processing with low cost and fast computational performance, even when large-capacity computing resources are required, such as for artificial intelligence (AI) computation. Background Technology
[0003] Thanks to advancements in computer-related technology and price reductions due to mass production, high-performance computer resources, particularly GPU resources, have become available and are being operated even in general personal computers (PCs). Additionally, thanks to advancements in network technology, each high-performance computer is connected to others through a network.
[0004] A cluster computer is a collection of multiple independent, stand-alone computer systems that are interconnected in parallel or serially while maintaining a distributed form. It is used (operated) as a single integrated computing resource, provides better performance and stability than a single computer, and is much more cost-effective than a single computer of a similar level.
[0005] Clusters and general distributed systems differ in the relationships between their parts; clusters have peer-to-peer relationships, and it is desirable to use independent networks.
[0006] Distributed systems use an underlying communication layer that is peer-to-peer, and there are no intrinsic hierarchies or other configurations; only a flat list of communicating entities exists. At a higher level of abstraction, communicating entities are generally organized into a client-server paradigm to simplify the complexity of the system.
[0007] Meanwhile, generally, each high-performance computer rarely uses 100% of its resources even during normal operation, and since the time of use is also very limited, a certain amount of idle computing resources (unused resources, Idle Computing Resource) is always generated.
[0008] By utilizing the idle computing resources of each high-performance computer connected via a network, it is possible to configure information processing capabilities equivalent to a supercomputer or higher-level computer, and this concept is referred to as grid computing.
[0009] In other words, a grid computer is a concept that connects distributed computing resources via a network to operate as a single computer, offering the advantages of not only utilizing idle computing resources but also fast data processing (task) speeds.
[0010] Nevertheless, each high-performance computer has a problem of wasting unused high-performance idle resources by failing to utilize them.
[0011] A prior art that partially resolves these needs and problems is the ‘system and method for managing idle computing resources’ according to Korean Patent Registration No. 10-1083628 (November 9, 2011).
[0013] FIG. 1 is a functional configuration diagram of a system for managing idle computing resources according to an embodiment of the prior art.
[0014] Referring to the attached drawings, a system configuration for managing idle computing resources according to an embodiment of the prior art is described in detail and includes a rental system (110 to 118), a network (100), an intermediary system (130), a rental system (150 to 158), and a measurement system (170).
[0015] The rental system (110 to 118) connects to the rental system (150 to 158) via the network (100) and utilizes the provided idle computing resources, wherein the minimum and maximum rental fees proposed by each are adjusted to an intermediate value reflecting the rental unit demand index measured by the measurement system (170) by the brokerage system (130).
[0016] The prior art has the advantage of enabling rapid transaction agreement by determining and presenting the optimal median value, which reflects the lease unit demand index measured by the brokerage system, between the maximum lease price proposed through autonomous competition and the minimum lease price proposed by the lease system for the idle computing resources of the lease system.
[0017] However, conventional rental systems fail to identify the specifications, processing capacity, and permitted time slots of idle computing GPU resources provided by the rental system. While they have the advantage of reaching quick agreements through transactions concluded at the median value of prices presented based on performance and autonomous competition, they have a problem in that they do not satisfy the specifications, processing capacity, and permitted time slots for high-performance computing GPU resources intended for artificial intelligence (AI) computation.
[0018] Therefore, in preparation for situations requiring high-performance, large-capacity GPU computing resources, such as artificial intelligence (AI) computations, it is necessary to develop a technology that periodically collects operational information—including GPU specifications, capacity, and operating time—of high-performance computers connected to a network in real time through an agent, and occupies unused idle computing GPU resources in real time at the required time with the necessary specifications and capacity. Prior art literature
[0020] Republic of Korea Patent Registration No. 10-1083628 (November 09, 2011) 'System and method for managing idle computing resources' The problem to be solved
[0021] The present invention, devised to resolve the problems and necessities of the conventional technology described above, provides a cluster computing system and a method of operation for artificial intelligence data processing comprising automatic detection of idle computing GPU resources, wherein operational information including the corresponding GPU specifications, processing capacity, and available time slots of each high-performance computer connected to a network is regularly collected in real time through an agent, recorded and managed as idle computing GPU resources, and idle computing resources with the specifications, capacity, and performance required by a GPU cluster processing large-scale artificial intelligence computations are rapidly built into a cluster at a low cost during the required time slot and provided in real time. means of solving the problem
[0023] A cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to the present invention, devised to achieve the above-mentioned purpose, comprises: a cluster terminal unit (100) that outputs data requiring artificial intelligence processing and a corresponding request signal based on a command signal input by a resource agent app that is downloaded, installed, and operated in an active state, and performs password authentication connection and encrypted communication with a counterparty; a cluster communication network unit (200) that connects to the cluster terminal unit (100), sets up switching to connect one or two communication paths with the called party based on a request signal from the calling party, and blocks the connection of the connected communication path based on a hook-off request signal from one of the connected parties; and a cluster server unit (300) that connects a portion of the idle computing GPU resources detected and recorded and managed based on a request signal from the cluster terminal unit (100) connected via the cluster communication network unit (200), operates it as a distributed artificial intelligence cluster for requested artificial intelligence data processing, performs password authentication connection and encrypted communication with a selected counterparty, and provides the latest version of the resource agent app. It may include a cluster resource providing terminal (400) that connects to the cluster server unit (300) via the cluster communication network unit (200), is installed and operated in an active state, and provides idle resource information including the specifications of the GPU resource provided by the cluster server unit (300), the video RAM capacity, and the driver version in response to the corresponding request signal of the cluster server unit (300).
[0024] It may further include a cluster idle resource database unit (500) that detects and collects idle resource information recorded in a standardized format via a corresponding control signal of the cluster server unit (300) connected via the cluster communication network unit (200), records it in an allocated memory area, and outputs it by searching.
[0025] The cluster terminal unit (100) comprises: a terminal control unit (110) that connects to each functional unit provided in the cluster terminal unit (100) and outputs corresponding control signals based on the built-in operating program, resource agent app program, data, and input command signals; a resource agent app operation unit (120) that connects to the terminal control unit (110), records and stores the latest version of the resource agent app program downloaded via streaming from the cluster server unit (300) based on the corresponding control signal in an allocated memory area, and automatically installs and operates the resource agent app based on the corresponding control signal remotely authorized from the cluster server unit (300); and a terminal password authentication communication unit (130) that connects to the cluster communication mode when a mutually authenticated state is confirmed by mutually exchanging connection passwords with the cluster server unit (300) based on the corresponding control signals authorized from the terminal control unit (110) and the resource agent app, respectively. It may include: a terminal frame encryption / decoding unit (140) that encrypts and converts a signal transmitted to the cluster server unit (300) connected in cluster communication mode by a corresponding control signal authorized by the terminal control unit (110) and the resource agent app, respectively, into a packet frame for transmission, and decrypts and converts a signal received as a packet frame for input; a terminal multi-channel occupancy communication unit (150) that connects to the terminal control unit (110), connects to the cluster communication network unit (200) by a corresponding control signal, requests the allocation of two communication paths for simultaneous connection, and transmits and receives communication signals of the same content in both directions; and a terminal touch input / output unit (160) that connects to the terminal control unit (110), outputs a multimedia signal output from the cluster terminal unit (100) by a corresponding control signal, and inputs an external touch signal and a multimedia signal.
[0026] The above terminal control unit (110) is characterized by being configured to calculate the signal-to-noise ratio values of the signals input from the terminal multi-channel occupancy communication unit (150) to two communication paths, output a corresponding control signal to input the signal received through the best communication path, and if analyzed as having the same value, output a corresponding control signal to input the signal received through any arbitrarily selected communication path.
[0027] The above cluster communication network unit (200) comprises: a cluster communication network management unit (210) that connects to each functional unit constituting the cluster communication network unit (200) and outputs and monitors corresponding control signals based on built-in programs and operational data; a mobile communication channel unit (220) that connects to the cluster communication network management unit (210), connects to the calling side and the called side respectively based on the corresponding control signal, transmits communication signals bidirectionally using a CDMA mobile communication method, and has one or more communication paths; and a mobile communication switching unit (230) that connects to the cluster communication network management unit (210), allocates one or more communication paths provided in the mobile communication channel unit (220) based on the corresponding control signal to connect the calling side and the called side, and sets the switching for call connection and call blocking of the connected communication paths. A Wi-Fi channel unit (240) connected to the cluster communication network management unit (210) and having one or more communication paths that are connected to the calling side and the called side respectively by a control signal and transmit communication signals bidirectionally using a Wi-Fi communication method; a Wi-Fi switching unit (250) connected to the cluster communication network management unit (210) and having one or two communication paths provided by the Wi-Fi channel unit (240) allocated by a control signal to connect the calling side and the called side and set switching for call connection and call blocking of the communication path; and a Bluetooth channel unit (260) connected to the cluster communication network management unit (210) and having one or more communication paths that are connected to the calling side and the called side respectively by a control signal and transmit communication signals bidirectionally using a Bluetooth communication method. A Bluetooth switching unit (270) that connects to the cluster communication network management unit (210) and, by means of a corresponding control signal, allocates one or two communication paths provided by the Bluetooth channel unit (260) to connect the calling party and the called party, and sets the switching of call connection and call blocking of the communication path;It may include: a wired channel switching unit (280) that connects to the cluster communication network management unit (210) and, by means of a control signal, allocates one or two wired communication paths to connect the calling party and the called party, and sets the switching of call connection and call blocking of the communication path that transmits a wired communication method communication signal in both directions; and a cluster network database unit (290) that connects to the cluster communication network management unit (210) and, by means of a control signal, records and stores data generated during operation in the cluster communication network unit (200) and command signals input from the outside in a time series in respective allocated memory areas.
[0028] The cluster server unit (300) is connected to each functional unit provided in the cluster server unit (300), and a server centralized operation unit (302) outputs a corresponding control signal to each functional unit based on a command signal input to an embedded and installed operation program and a resource agent app program; a server agent app management unit (304) records and manages resource agent app programs by version based on the corresponding control signal of the server centralized operation unit (302), monitors that the latest version is installed and operated at the connected cluster terminal unit (100) or cluster resource provision terminal unit (400), and outputs a corresponding control signal to set the operation status by downloading and automatically installing via a streaming method; and a server password authentication communication unit (306) that connects to the cluster terminal unit (100) or cluster resource provision terminal unit (400) in cluster communication mode when a mutually authenticated state is confirmed by mutually exchanging connection passwords based on the corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app. A server frame encryption / decoding unit (308) that encrypts and converts a signal transmitted to the cluster terminal (100) or cluster resource provision terminal (400) connected in cluster communication mode by a corresponding control signal authorized by the server central operation unit (302) and the resource agent app, respectively, into a packet frame for transmission, and decrypts and converts a signal received as a packet frame for input; and an idle GPU resource detection unit (310) that connects to the cluster resource provision terminal (400) in cluster communication mode by a corresponding control signal authorized by the server central operation unit (302) and the resource agent app, respectively, and detects an idle computing resource equipped with a GPU. A resource metadata collection unit (312) that collects metadata of GPU resources provided in idle computing resources detected by the idle GPU resource detection unit (310) by means of corresponding control signals authorized respectively from the server centralized operation unit (302) and the resource agent app;A metadata standardization unit (314) that records and manages metadata collected by the resource metadata collection unit (312) by means of corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, respectively, by entering it into a standardized form; and a cluster management unit (316) that analyzes data requesting artificial intelligence processing from the cluster terminal unit (100) by means of corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, respectively, estimates the required number and specifications of GPUs, video memory capacity, model name, and driver version, and connects the cluster resource provision terminal unit (400) of idle computing resources corresponding to a multiple of 1.5 in a cluster communication mode state to form and manage a distributed artificial intelligence cluster. A cluster network traffic analysis unit (318) that estimates the transmission error rate, transmission delay, and physical distance by transmitting and receiving test data on each communication path allocated to connect each cluster resource providing terminal (400) constituting the distributed artificial cluster formed by the cluster management unit (316) by means of corresponding control signals authorized respectively from the server centralized operation unit (302) and the resource agent app; and an orchestration management unit (320) that organically integrates each GPU resource of each cluster resource providing terminal (400) constituting the distributed artificial cluster connected by the cluster management unit (316) by means of corresponding control signals authorized respectively from the server centralized operation unit (302) and the resource agent app, and coordinates mutual data processing functions to eliminate the occurrence of bottlenecks in data processing. It may include a server multi-channel occupancy communication unit (322) that connects to the cluster communication network unit (200) by means of corresponding control signals authorized respectively from the server centralized operation unit (302) and the resource agent app, requests the allocation of two communication paths to connect simultaneously, and transmits and receives communication signals of the same content in both directions.;
[0029] The above-mentioned server centralized operation unit (302), server agent app management unit (304), server password authentication communication unit (306), server frame encryption / decryption unit (308), idle GPU resource detection unit (310), resource metadata collection unit (312), metadata standardization unit (314), cluster management unit (316), cluster network traffic analysis unit (318), orchestration management unit (320), and server multi-channel occupancy communication unit (322) are each made of a printed circuit board (PCB) and are mounted in the form of a server rack on the top of a slot (SL) installed in a state corresponding to the inner left and right walls of the module enclosure unit (600), and are connected to a main board (B) fixed to the rear of the module enclosure unit (600) and generate heat during operation, and an airflow generator (610) is installed on each side to air-cool the heat state, and the above An airflow generator (610) is installed on the exterior of each of the two sides of the module housing (600), and the two sides of the module housing (600) are provided with a plurality of exhaust holes (620) formed at an upward inclination, and an upward airflow induction path (630) is formed between the airflow generator (610) and the two sides of the module housing (600), and a blower (640) is installed at the bottom of the airflow generator (610) to supply air to create an upward airflow through the upward airflow induction path (630), wherein the upward airflow induction path (630) is characterized by having a shape in which the path space widens as it goes upward.
[0030] The above module housing (600) and airflow generator (610) are spray-coated with a surface coating agent to provide waterproofness, anti-contamination properties, and durability, and the surface coating agent is characterized by a composition comprising, for every 100 parts by weight of transparent polypropylene resin, 5 parts by weight of fine mica powder, 10 parts by weight of methylsulfonic methane, 3.5 parts by weight of sodium boroshydride, 15 parts by weight of polytetrafluoroethylene, 10 parts by weight of urea, 10 parts by weight of phosphite, 5 parts by weight of gluconate, and 10 parts by weight of sodium bicarbonate.
[0032] The method for operating a cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing, devised to achieve the above-mentioned purpose, comprises: a cluster terminal unit, a cluster communication network unit, a cluster server unit, a cluster resource providing terminal unit, and a cluster idle resource database unit; wherein, when it is determined that an operation command signal is input by the cluster server unit, the operation data is initialized and the latest version of the resource agent app is installed on the cluster terminal unit and the cluster resource providing terminal unit connected to the cluster communication network unit; a resource detection process in which a cluster communication mode is set by the cluster server unit of the preparation process, a cluster resource providing terminal unit is connected to, and an idle GPU resource detection routine is operated to detect idle GPU resources; and a metadata collection process in which an idle resource metadata collection routine is operated to collect corresponding GPU information from the idle GPU resources detected by the cluster server unit of the resource detection process. It may include: a standardization process that operates a metadata standardization conversion routine to record information collected by the cluster server unit of the above metadata collection process in a standardized format; a cluster process that operates a cluster management routine when it is determined that a command signal for artificial intelligence data processing is input by the cluster server unit of the above standardization process; a traffic process that operates a traffic environment evaluation routine by the cluster server unit of the above cluster process; and an orchestration process that operates an orchestration routine by the cluster server unit of the above traffic process.
[0033] The above traffic process may further include a feedback process in which, if it is determined by the cluster server part of the above traffic process that the artificial intelligence data processing is not completed (S190), feedback is given to the cluster process (S160), and if it is determined that the artificial intelligence data processing is completed, a repeating command signal is input (S200), and if it is determined that the repeating command signal is not input, the process proceeds to termination.
[0034] The above idle GPU resource detection routine may include: an idle resource detection first step of detecting an idle GPU resource from a cluster resource providing terminal connected by the cluster server unit and setting a cluster communication mode; an idle resource detection second step of determining whether an idle GPU resource target has been detected by the cluster server unit of the idle resource detection first step; and an idle resource detection third step of selecting a cluster resource providing terminal determined to have been detected by the cluster server unit of the idle resource detection second step as an idle GPU resource, verifying location information including the corresponding description information and IP address, and recording and managing it in an allocated memory area.
[0035] The above idle resource metadata collection routine may include: a first step of resource metadata collection in which the cluster server unit confirms and connects to the location information of the cluster resource providing terminal unit selected as an idle GPU resource during the resource detection process; a second step of resource metadata collection in which the cluster server unit of the first step of resource metadata collection sets a cluster communication mode with the cluster resource providing terminal unit selected as an idle GPU resource; and a third step of resource metadata collection in which the cluster server unit of the second step of resource metadata collection collects the metadata of the idle GPU resource connected in the cluster communication mode and records it in an allocated memory area in a state associated with the corresponding location information.
[0036] The above metadata standardization conversion routine may include: a first standardization conversion step in which idle GPU resource information of a cluster resource providing terminal, collected and recorded during the metadata collection process by the cluster server, is read and converted into character data recognizable by a large-scale language model of artificial intelligence and a natural language processing program; and a second standardization conversion step in which the character data converted by the cluster server of the first standardization conversion step is recorded and managed in corresponding areas of a standardized format.
[0037] The above cluster management routine may include: a first cluster management step by the cluster server unit analyzing the scale of idle GPU resources required to process input data for artificial intelligence; a second cluster management step by the cluster server unit of the first cluster management step determining the specifications, video RAM capacity, model, and version of the GPU to configure the distributed artificial intelligence cluster of the analyzed processing scale, and estimating the required quantity of idle GPU resources consisting of cluster resource providing terminals; and a third cluster management step by the cluster server unit of the second cluster management step selecting idle GPU resources corresponding to 1.5 times the estimated required quantity and designating them as cluster groups.
[0038] The above traffic environment evaluation routine may include: a first step of traffic environment evaluation in which the cluster server unit requests and connects to a communication path allocated to each cluster resource providing terminal unit constituting a cluster group designated in the cluster process and the cluster communication network unit; a second step of traffic environment evaluation in which test data is transmitted to the cluster resource providing terminal unit connected by the cluster server unit of the first step of traffic environment evaluation and a received signal is input; a third step of traffic environment evaluation in which the cluster server unit of the second step of traffic environment evaluation analyzes the signal-to-noise ratio of the input signal to measure the data transmission error rate of the corresponding communication path; a fourth step of traffic environment evaluation in which the cluster server unit of the third step of traffic environment evaluation analyzes the delay time of the input signal to measure the connection delay time of the corresponding communication path; and a fifth step of traffic environment evaluation in which the cluster server unit of the fourth step of traffic environment evaluation analyzes the node information through which the input signal has passed to measure the physical distance of the corresponding communication path.
[0039] The above orchestration routine may include: a first orchestration stage in which the cluster server unit analyzes values measured in the traffic process and sequentially sorts idle GPU resources in order from lowest to highest, starting from the idle GPU resources with excellent traffic environments; a second orchestration stage in which the cluster server unit of the first orchestration stage connects idle GPU resources corresponding to one multiple of the required quantity from the upper order of the sequentially sorted idle GPU resources via a network to form a distributed artificial intelligence cluster and utilizes it for artificial intelligence data processing; and a third orchestration stage in which the cluster server unit of the second orchestration stage operates the distributed artificial intelligence cluster connected via a network to operate a computational orchestration subroutine.
[0040] The above computational orchestration subroutine may include: a first computational orchestration stage in which the cluster server unit converts application programs, each operated at each cluster resource provider terminal connected to the distributed artificial cluster, into containerized programs by Kubernetes processing so that they can all be used in common; a second computational orchestration stage in which the cluster server unit of the first computational orchestration stage schedules the allocation of resources by slump processing so that each application program can continuously use resources; and a third computational orchestration stage in which the cluster server unit of the second computational orchestration stage processes data transmitted and received by each application program via MPI to ensure consistent data communication. Effects of the invention
[0042] The present invention, configured as described above, has the advantage of regularly and continuously collecting the corresponding GPU specifications and operational information of each high-performance computer connected to a network in real time by an agent, recording and managing them as idle computing GPU resources, and rapidly building and providing a cluster composed of idle computing GPU resources with the specifications, capacity, and performance required for large-scale artificial intelligence computation processing in real time at a low cost at the required time. Brief explanation of the drawing
[0044] FIG. 1 is a functional configuration diagram of a system managing idle computing resources according to an embodiment of the prior art; FIG. 2 is a functional configuration diagram of a cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to an embodiment of the present invention; FIG. 3 is a detailed functional configuration diagram of a cluster terminal unit according to an embodiment of the present invention; FIG. 4 is a detailed functional configuration diagram of a cluster communication network unit according to an embodiment of the present invention; FIG. 5 is a detailed functional configuration diagram of a cluster server unit according to an embodiment of the present invention; FIG. 6 is a detailed functional configuration diagram of a module enclosure unit according to an embodiment of the present invention; FIG. 7 is a signal flow diagram of a cluster computing system operation method configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to an embodiment of the present invention; FIG. 8 is a signal flow diagram of an idle GPU resource detection routine according to an embodiment of the present invention; FIG. 9 is a signal flow diagram of an idle resource metadata collection routine according to an embodiment of the present invention; FIG. 10 is an FIG. 11 is a signal flow diagram of a metadata standardization conversion routine according to an embodiment of the present invention, FIG. 12 is a signal flow diagram of a traffic environment evaluation routine according to an embodiment of the present invention, FIG. 13 is a signal flow diagram of an orchestration routine according to an embodiment of the present invention, and FIG. 14 is a signal flow diagram of an operation orchestration subroutine according to an embodiment of the present invention. Specific details for implementing the invention
[0045] The present invention is capable of various modifications and may have various embodiments; specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the present invention. In describing the present invention, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions may obscure the essence of the present invention.
[0046] In the following, "connection" and "contact" refer to the state of signal transmission and reception and have the same meaning; they will be used selectively depending on the context.
[0048] FIG. 2 is a functional configuration diagram of a cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to an embodiment of the present invention, FIG. 3 is a detailed functional configuration diagram of a cluster terminal unit according to an embodiment of the present invention, FIG. 4 is a detailed functional configuration diagram of a cluster communication network unit according to an embodiment of the present invention, FIG. 5 is a detailed functional configuration diagram of a cluster server unit according to an embodiment of the present invention, and FIG. 6 is a detailed functional configuration diagram of a module enclosure unit according to an embodiment of the present invention.
[0049] Hereinafter, a cluster computing system (90) configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to an embodiment of the present invention will be described in detail with reference to all attached drawings. The system further comprises a cluster terminal unit (100), a cluster communication network unit (200), a cluster server unit (300), a cluster resource provision terminal unit (400), and a cluster idle resource database unit (500).
[0050] The cluster terminal (100) outputs data requiring artificial intelligence processing and a corresponding request signal based on a command signal input from a resource agent app that is downloaded, installed, and operated in an active state, and establishes a password authentication connection with the other party and communicates encrypted.
[0051] The cluster terminal unit (100) is configured to include a terminal control unit (110), a resource agent app operation unit (120), a terminal password authentication communication unit (130), a terminal frame encryption / decryption unit (140), a terminal multi-channel occupancy communication unit (150), and a terminal touch motion sickness input / output unit (160).
[0052] The terminal control unit (110) connects to each functional unit provided in the cluster terminal unit (100) and outputs a corresponding control signal based on the built-in operating program, resource agent app program, data, and input command signal.
[0053] The resource agent app operation unit (120) connects to the terminal control unit (110) and records and stores the latest version of the resource agent app program, which is downloaded via streaming from the cluster server unit (300) by the corresponding control signal, in an allocated memory area, and automatically installs and operates it by the corresponding control signal that is remotely authorized from the cluster server unit (300).
[0054] The terminal password authentication communication unit (130) exchanges a connection password with the cluster server unit (300) by means of the corresponding control signal authorized by the terminal control unit (110) and the resource agent app, respectively, and connects to the cluster communication mode when the mutually authenticated state is confirmed.
[0055] The terminal encryption authentication communication unit (130) uses a communication technology that presents mutual encryption keys in the SSL / TLS manner and connects to transmit communication signals to each other when verified. If the SSL / TLS encryption key is not presented and verified by the terminal encryption authentication communication unit (130), the system is set to a state where it cannot make contact with the other computer to transmit, receive, or process data. The terminal encryption authentication communication unit (130) has the advantage of blocking unauthorized access by others.
[0056] The terminal frame encryption / decryption unit (140) encrypts and converts signals transmitted to the cluster server unit (300) connected in cluster communication mode by the respective control signals authorized from the terminal control unit (110) and the resource agent app, and transmits the signals received encrypted as packet frames by decrypting and converting them into input.
[0057] Here, the packet frame consists of a total of four fields, for example, a start field consisting of 10 words, a data field consisting of 50 words, a check field consisting of 20 words, and an end field consisting of 10 words, and 1 word consists of 8 bytes, and 1 byte consists of 8 bits. The number of fields, the number of bytes constituting the words, and the number of bits constituting the bytes can be increased or decreased as needed, and in this case, there is an advantage of securely encrypting communication against unauthorized hacking, tampering, etc., by a third party who does not know the respective constituent numbers.
[0058] The terminal multi-channel occupancy communication unit (150) connects to the terminal control unit (110) and connects to the cluster communication network unit (200) by means of the corresponding control signal, requests the allocation of two communication paths to connect simultaneously, and transmits and receives communication signals of the same content in both directions.
[0059] The terminal control unit (110) calculates the signal-to-noise ratio (SNR) value of the signal input to two communication paths from the terminal multi-channel occupancy communication unit (150) and outputs a control signal to input the signal received through the best communication path, and if analyzed as having the same value, outputs a control signal to input the signal received through any arbitrarily selected communication path.
[0060] The terminal touch input / output unit (160) is connected to the terminal control unit (110) and outputs a multimedia signal output from the cluster terminal unit (100) according to the corresponding control signal, and inputs an external touch signal and a multimedia signal.
[0061] The cluster communication network unit (200) connects to the cluster terminal unit (100) and sets up a switching connection to one or two communication paths with the called party based on a request signal from the calling party, and blocks the connection of the connected communication path based on a hook-off request signal from either of the connected parties.
[0062] The cluster communication network unit (200) is configured to include a cluster communication network management unit (210), a mobile communication channel unit (220), a mobile communication switching unit (230), a Wi-Fi channel unit (240), a Wi-Fi switching unit (250), a Bluetooth channel unit (260), a Bluetooth switching unit (270), a wired channel switching unit (280), and a cluster network DB unit (290).
[0063] The cluster communication network management unit (210) connects to each functional unit constituting the cluster communication network unit (200) and outputs and monitors the corresponding control signals based on the built-in programs and operation data.
[0064] The mobile communication channel unit (220) is connected to the cluster communication network management unit (210) and is connected to the calling side and the called side respectively by the corresponding control signal, and transmits communication signals bidirectionally using the CDMA mobile communication method and is provided with one or more communication paths.
[0065] The mobile communication switching unit (230) connects to the cluster communication network management unit (210) and, by means of the corresponding control signal, allocates one or more communication paths provided in the mobile communication channel unit (220) to connect the calling party and the called party, and switches the call connection and call blocking of the connected communication path.
[0066] The Wi-Fi channel unit (240) is connected to the cluster communication network management unit (210) and is connected to the calling side and the called side respectively by the corresponding control signal, and has one or more communication paths that transmit communication signals bidirectionally using the Wi-Fi communication method.
[0067] The Wi-Fi switching unit (250) connects to the cluster communication network management unit (210) and, by means of the corresponding control signal, allocates one or two communication paths provided by the Wi-Fi channel unit (240) to connect the calling party and the called party, and sets the switching of call connection and call blocking of the communication path.
[0068] The Bluetooth channel unit (260) is connected to the cluster communication network management unit (210) and, by means of the corresponding control signal, is connected to the calling side and the called side respectively and has one or more communication paths that transmit communication signals bidirectionally using the Bluetooth communication method.
[0069] The Bluetooth switching unit (270) connects to the cluster communication network management unit (210) and, by means of the corresponding control signal, allocates one or two communication paths provided by the Bluetooth channel unit (260) to connect the calling party and the called party, and sets the switching of call connection and call blocking of the communication path.
[0070] The wired channel switching unit (280) is connected to the cluster communication network management unit (210) and, by means of a control signal, allocates one or two wired communication paths to connect the calling party and the called party, and switches the call connection and call blocking of the communication paths that transmit communication signals of the wired communication method in both directions.
[0071] The cluster network database unit (290) connects to the cluster communication network management unit (210) and, by means of the corresponding control signal, records and stores data generated during operation in the cluster communication network unit (200) and command signals input from the outside in a time series in the respective allocated memory areas.
[0072] The cluster server unit (300) connects a portion of idle computing GPU resources that are detected and managed by a request signal from a cluster terminal unit (100) connected via the cluster communication network unit (200), operates it as a distributed artificial intelligence cluster for processing requested artificial intelligence data, establishes a password authentication connection with a selected counterparty, communicates encrypted, and provides the latest version of the resource agent app.
[0073] Here, a distributed artificial cluster is defined and understood as a cluster computer in the same sense as a single large computer or supercomputer, where multiple independent computers are connected via a network to perform the same function.
[0074] The cluster server unit (300) is configured to include a server centralized operation unit (302), a server agent app management unit (304), a server password authentication communication unit (306), a server frame encryption / decryption unit (308), an idle GPU resource detection unit (310), a resource metadata collection unit (312), a metadata standardization unit (314), a cluster management unit (316), a cluster network traffic analysis unit (318), an orchestration management unit (320), and a server multi-channel occupancy communication unit (322).
[0075] The server centralized operation unit (302) connects to each functional unit equipped in the cluster server unit (300) and outputs a corresponding control signal to each functional unit based on the built-in and installed operation program and the resource agent app program and the input command signal.
[0076] The server agent app management unit (304) records and manages resource agent app programs by version according to the corresponding control signal of the server centralized operation unit (302), monitors that the latest version is installed and operated at the connected cluster terminal unit (100) or cluster resource provision terminal unit (400), and outputs the corresponding control signal to set the operating status by downloading and automatically installing via a streaming method.
[0077] The server password authentication communication unit (306) exchanges a connection password with the cluster terminal unit (100) or the cluster resource provision terminal unit (400) by means of the corresponding control signal authorized by the server central operation unit (302) and the resource agent app, respectively, and connects to the cluster communication mode when the mutually authenticated state is confirmed.
[0078] The server frame encryption / decryption unit (308) encrypts and converts signals transmitted to the cluster terminal unit (100) or cluster resource provision terminal unit (400) connected in cluster communication mode by the respective control signals authorized from the server centralized operation unit (302) and the resource agent app, respectively, into packet frames for transmission, and decrypts and converts signals received as packet frames for input.
[0079] The idle GPU resource detection unit (310) connects to the cluster resource provision terminal (400) in cluster communication mode by means of the corresponding control signals authorized by the server central operation unit (302) and the resource agent app, respectively, and detects idle computing resources equipped with GPUs.
[0080] The resource metadata collection unit (312) collects metadata of GPU resources provided in idle computing resources detected by the idle GPU resource detection unit (310) by means of corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, respectively.
[0081] The metadata standardization unit (314) records and manages the metadata collected by the resource metadata collection unit (312) by entering it into a standardized form based on the corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, respectively.
[0082] The cluster management unit (316) analyzes data requesting artificial intelligence processing from the cluster terminal unit (100) by means of the corresponding control signals authorized by the server central operation unit (302) and the resource agent app, respectively, estimates the required number and specifications of GPUs, video memory capacity, model name and driver version, and connects the cluster resource provision terminal unit (400) of idle computing resources corresponding to 1.5 times the number in a cluster communication mode state to form and manage a distributed artificial intelligence cluster.
[0083] Here, one GPU is described as meaning one computer (PC), and if necessary, two or more GPUs may be provided in one computer.
[0084] The cluster network traffic analysis unit (318) transmits and receives test data to each communication path allocated to connect each cluster resource providing terminal (400) that constitutes the distributed artificial cluster formed by the cluster management unit (316) by means of the respective control signals authorized from the server central operation unit (302) and the resource agent app, and estimates the transmission error rate, transmission delay, and physical distance.
[0085] The orchestration management unit (320) organically integrates the GPU resources of each cluster resource providing terminal (400) that constitutes the distributed artificial cluster connected by the cluster management unit (316) by means of the corresponding control signals authorized by the server central operation unit (302) and the resource agent app, and coordinates mutual data processing functions to eliminate the occurrence of bottlenecks in data processing. The orchestration management unit (320) is operated by a program that includes Kubernetes, Slurm, and MPI (Message Passing Interface).
[0086] The server multi-channel occupancy communication unit (322) connects to the cluster communication network unit (200) by means of the corresponding control signals authorized by the server central operation unit (302) and the resource agent app, respectively, requests the allocation of two communication paths to connect simultaneously, and transmits and receives communication signals of the same content in both directions.
[0087] The cluster resource provision terminal (400) connects to the cluster server unit (300) via the cluster communication network unit (200) and provides idle resource information including the specifications of the GPU resource, video RAM capacity, and driver version provided by the cluster server unit (300) and the corresponding request signal of the resource agent app installed and operating in an active state.
[0088] The cluster idle resource database unit (500) detects and collects idle resource information in a standardized format via the cluster communication network unit (200) and records it in an allocated memory area, and outputs it by searching, based on the corresponding control signal of the connected cluster server unit (300).
[0089] The server centralized operation unit (302), server agent app management unit (304), server password authentication communication unit (306), server frame encryption / decryption unit (308), idle GPU resource detection unit (310), resource metadata collection unit (312), metadata standardization unit (314), cluster management unit (316), cluster network traffic analysis unit (318), orchestration management unit (320), and server multi-channel occupancy communication unit (322) are each made of printed circuit boards (PCB) and are mounted in the form of a server rack on the top of the slot (SL) and the slot (SL) installed in a state corresponding to the inner left wall and right wall of the module housing (600), and are connected to the main board (B) fixed to the rear of the module housing (600) and generate heat during operation, and an airflow generator (610) is installed on the side to air cool the heat.
[0090] An airflow generator (610) is installed on the exterior of each side of the module housing (600), and a plurality of exhaust holes (620) are formed by perforating them at an upward angle on both sides of the module housing (600). An upward airflow inducing channel (630) is formed between the airflow generator (610) and both sides of the module housing (600), and a blower (640) is installed at the bottom of the airflow generator (610) to supply air to create an upward airflow through the upward airflow inducing channel (630). The upward airflow inducing channel (630) has a shape in which the channel space widens as it goes upward.
[0091] The reason for forming it in this way is to create a rapid updraft by applying high pressure from below, and to allow it to spread outward to the upper side with slightly lower pressure from above.
[0092] The exhaust port (620) allows the air inside the module housing (600) to be naturally discharged to the outside through a kind of venturi effect, thereby providing a cooling effect.
[0093] A blower (640) is installed at the bottom of the airflow generator (610) to generate airflow.
[0094] Through this configuration, natural exhaust is possible without external air flowing into the module housing (600), thereby increasing the cooling efficiency inside the module housing (600).
[0095] The module housing (600) and the airflow generator (610) are spray-coated with a surface coating agent to provide water resistance, anti-contamination properties, and durability. The surface coating agent is composed of 5 parts by weight of fine mica powder, 10 parts by weight of methylsulfonic methane, 3.5 parts by weight of sodium boroshydride, 15 parts by weight of polytetrafluoroethylene, 10 parts by weight of urea, 10 parts by weight of phosphite, 5 parts by weight of gluconate, and 10 parts by weight of sodium bicarbonate, mixed with 100 parts by weight of transparent polypropylene resin.
[0096] Polypropylene resin is a hydrocarbon chemically composed only of carbon and hydrogen, so it gives a smooth, slippery appearance and texture similar to candles or soap. It has a specific gravity of 0.91, making it lighter than water and allowing it to float. It is widely used in industrial applications due to its excellent strength and chemical resistance, and it is characterized by high strength, lightness, and the fact that it does not absorb moisture at all.
[0097] Sodium bicarbonate enhances durability, antifouling properties, and water resistance while increasing moldability due to the high crosslinking density of polypropylene resin.
[0098] Phosphites enhance chemical resistance and heat insulation, while contributing to the suppression of deformation due to improved strength, hardness, and wear resistance, and gluconates prevent interfacial separation to improve fixation and suppress surface slipperiness.
[0099] Urea enhances moisture resistance by securing heat resistance through the formation of a eutectic point.
[0100] Polytetrafluoroethylene enhances thermal insulation properties by securing chemical resistance, electrical insulation properties, non-stick properties, anti-fouling properties, heat resistance, and friction properties, thereby strengthening the inhibition of heat conduction.
[0101] Sodium borohydride increases bond strength to ensure heat resistance, while enhancing water repellency, moisture resistance, water resistance, and antifouling properties. Methylsulfonic methane is added to suppress cracking and splitting while maintaining flexibility.
[0102] Fine mica powder is a representative heat-insulating agent.
[0103] Therefore, the module housing (600) and the airflow generator (610) have the advantage of maintaining excellent performance in waterproofing, anti-contamination, and durability for a long period of time, and extending their lifespan.
[0105] An "App" is a general term for application software. Since they are developed to be suitable for application in various fields, it is common for the developer to assign a name of their choice; however, if necessary, a name reflecting the application field may be assigned. When the size of the configured data is relatively small, it is also referred to as an "applet."
[0106] Both the CPU (Central Processing Unit) and the GPU (Graphic Processing Unit) are core components responsible for computation in computers, but they differ in their designed purposes and structures.
[0107] The CPU acts as the 'brain' of the computer, performs general operations and control, and rapidly switches and processes various tasks (general-purpose processing). The cores typically consist of 4 to 16 powerful cores (advantageous for high-performance single operations) and operate in a sequential (serial) processing manner.
[0108] GPUs perform graphics processing and large-scale parallel computations. They execute a very large number of identical operations simultaneously (parallel processing) and feature a simple structure consisting of hundreds to thousands of cores (advantageous for massive computation). They are suitable for graphics rendering, AI learning, image processing, and scientific computing, and operate using a parallel processing method. A GPU (Graphics Processing Unit) is an electronic circuit that accelerates digital image processing and computer graphics rendering. Because it processes large amounts of data simultaneously through a parallel structure, it is advantageous for tasks involving the repetition of identical or similar operations. It is also utilized in non-graphics domains that include artificial intelligence (AI), scientific computing, high-performance computing, and image processing. Factors to consider include video memory (VRAM), computational performance, architecture, and power efficiency.
[0109] In other words, the CPU is like a smart brain processing one task quickly and accurately at a time, for example, like one person calculating quickly and processing sequentially, whereas the GPU is like a powerful team of many people dividing similar tasks and processing them simultaneously, for example, like 1,000 people dividing and calculating simultaneously, so it processes much faster.
[0110] An orchestration system is a program that goes beyond simply processing individual cognitive functions in parallel to organically integrating them and coordinating their interactions to maximize the efficiency and adaptability of the entire system.
[0111] Orchestration systems include programs such as Kubernetes, Slurm, and MPI (Message Passing Interface).
[0112] Kubernetes is an open-source platform for deploying, managing, and scaling containerized applications (virtual boxes that ensure an application always behaves the same way regardless of where it is run), enabling efficient management and scaling of applications. Kubernetes enables: 1) automatic management of application updates through automated deployment and rollout, and rollback to previous versions in case of issues; 2) increasing or decreasing application instances when traffic changes are detected through auto-scaling; 3) automatic restarting or replacement of containers if they fail through self-healing; 4) increased application availability by distributing traffic across multiple container instances through load balancing; and 5) easy setup and management of communication between containers through service discovery.
[0113] Slurm is an open-source workload manager that schedules tasks to efficiently manage jobs and allocate resources in large computer clusters.
[0114] MPI (Message Passing Interface) is an interface (protocol) that standardizes inter-process communication in a parallel computing environment. It is a communication method that enables data exchange between HPC (High Performance Computing) systems using distributed memory structures, such as clusters or supercomputers, where multiple computers (or CPU cores) process (compute) a single program together.
[0116] FIG. 7 is a signal flow diagram of a cluster computing system operation method configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to an embodiment of the present invention, FIG. 8 is a signal flow diagram of an idle GPU resource detection routine according to an embodiment of the present invention, FIG. 9 is a signal flow diagram of an idle resource metadata collection routine according to an embodiment of the present invention, FIG. 10 is a signal flow diagram of a metadata standardization conversion routine according to an embodiment of the present invention, FIG. 11 is a signal flow diagram of a cluster management routine according to an embodiment of the present invention, FIG. 12 is a signal flow diagram of a traffic environment evaluation routine according to an embodiment of the present invention, FIG. 13 is a signal flow diagram of an orchestration routine according to an embodiment of the present invention, FIG. 14 is a signal flow diagram of a computation orchestration subroutine according to an embodiment of the present invention.
[0117] Hereinafter, with reference to all attached drawings, a method for operating a cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing according to an embodiment of the present invention will be described in detail. The method for operating a cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing comprises a cluster terminal, a cluster communication network, a cluster server, a cluster resource providing terminal, and a cluster idle resource database, and includes a preparation process, a resource detection process, a metadata collection process, a standardization process, a cluster process, a traffic process, an orchestration process, and a feedback process.
[0118] In the preparation process, when it is determined that an operation command signal is input by the cluster server (S100), the operation data is initialized, and the latest version of the resource agent app is installed on the cluster terminal and the cluster resource provider terminal connected to the cluster communication network (S110). The resource agent app is a dedicated application program used in the present invention.
[0119] The resource detection process sets the cluster communication mode by the cluster server part of the preparation process, connects to the cluster resource provision terminal, and operates an idle GPU resource detection routine to detect idle GPU resources (S120).
[0120] The idle GPU resource detection routine includes idle resource detection step 1, idle resource detection step 2, and idle resource detection step 3.
[0121] The first step of idle resource detection is to detect the target of idle GPU resources from the connected cluster resource providing terminal by connecting to the cluster communication network unit via the cluster server unit and setting the cluster communication mode (S121).
[0122] The second stage of idle resource detection determines whether the target of the idle GPU resource is detected by the cluster server part of the first stage of idle resource detection (S121) (S122).
[0123] The third step of idle resource detection selects a cluster resource providing terminal that is determined to have been detected by the cluster server part of the second step of idle resource detection (S122) as an idle GPU resource, verifies location information including the corresponding description information and IP address, and records and manages it in an allocated memory area (S123).
[0124] The metadata collection process operates an idle resource metadata collection routine that collects the corresponding GPU information from the idle GPU resources detected by the cluster server part of the resource detection process (S130).
[0125] The idle resource metadata collection routine includes resource metadata collection step 1, resource metadata collection step 2, and resource metadata collection step 3.
[0126] The first step of resource metadata collection is to verify the location information of the cluster resource providing terminal selected as an idle GPU resource during the resource detection process by the cluster server and connect to it (S131).
[0127] The second step of resource metadata collection sets the cluster communication mode with the cluster resource providing terminal selected as an idle GPU resource by the cluster server part of the first step of resource metadata collection (S131) (S132).
[0128] The third step of resource metadata collection involves collecting metadata of idle GPU resources connected in cluster communication mode by the cluster server unit of the second step of resource metadata collection (S132) and recording it in an allocated memory area in a state associated with the corresponding location information (S133).
[0129] The standardization process operates a metadata standardization conversion routine to record the information collected by the cluster server of the metadata collection process in a standardized format (S140).
[0130] The metadata structuring transformation routine is a configuration that includes the first structuring transformation stage and the second structuring transformation stage.
[0131] The first step of standardization conversion reads idle GPU resource information from the cluster resource providing terminal, which is collected and recorded during the metadata collection process by the cluster server, and converts it into character data that can be recognized by the artificial intelligence's large-scale language model and natural language processing program (S141).
[0132] The second step of standardization conversion records and manages the character data converted by the cluster server unit of the first step of standardization conversion (S141) in the corresponding area of the standardized format (S142).
[0133] When it is determined that a command signal for artificial intelligence data processing is input by the cluster server of the standardization process, the cluster process operates a cluster management routine (S160).
[0134] The cluster management routine includes cluster management step 1, cluster management step 2, and cluster management step 3.
[0135] The first step of cluster management is performed by the cluster server unit to analyze the scale of idle GPU resources required to process the input data into artificial intelligence (S161).
[0136] The second stage of cluster management determines the specifications, video RAM capacity, model, and version of the GPUs for configuring the distributed artificial cluster of the processing scale analyzed by the cluster server unit of the first stage of cluster management (S161), and estimates the required quantity of idle GPU resources consisting of the cluster resource provision terminal unit (S162).
[0137] The third stage of cluster management selects idle GPU resources corresponding to 1.5 times the required quantity estimated by the cluster server section of the second stage of cluster management (S162) and designates them as a cluster group (S163).
[0138] The traffic process operates a traffic environment evaluation routine by the cluster server part of the cluster process (S170).
[0139] The traffic environment assessment routine includes traffic environment assessment stage 1, traffic environment assessment stage 2, traffic environment assessment stage 3, traffic environment assessment stage 4, and traffic environment assessment stage 5.
[0140] The first step of the traffic environment evaluation is to request each cluster resource providing terminal and cluster communication network constituting the cluster group designated in the cluster process by the cluster server unit and connect to the allocated communication path (S171).
[0141] The second step of the traffic environment evaluation transmits test data to the cluster resource providing terminal connected by the cluster server part of the first step of the traffic environment evaluation (S171) and inputs the received signal (S172).
[0142] The third step of the traffic environment evaluation analyzes the signal-to-noise ratio of the signal input by the cluster server unit of the second step of the traffic environment evaluation (S172) to measure the data transmission error rate of the corresponding communication path (S173).
[0143] The fourth step of the traffic environment evaluation analyzes the delay time of the signal input by the cluster server unit of the third step of the traffic environment evaluation (S173) and measures the connection delay time of the corresponding communication path (S174).
[0144] The 5th step of the traffic environment evaluation analyzes the node information through which the signal input by the cluster server part of the 4th step of the traffic environment evaluation (S174) has passed, and measures the physical distance of the corresponding communication path (S175).
[0145] The orchestration process operates the orchestration routine by the cluster server part of the traffic process (S180).
[0146] The orchestration routine includes orchestration stage 1, orchestration stage 2, and orchestration stage 3.
[0147] The first stage of orchestration analyzes the values measured during the traffic process by the cluster server and sequentially sorts the idle GPU resources in order from lowest to highest, starting with the idle GPU resources with the best traffic environment (S181).
[0148] The second orchestration stage forms a distributed artificial intelligence cluster by connecting idle GPU resources corresponding to one multiple of the required quantity from the top order of idle GPU resources sequentially sorted by the cluster server part of the first orchestration stage (S181) via a network and utilizing them for artificial intelligence data processing (S182).
[0149] The third stage of orchestration operates a distributed artificial cluster connected via a network by the cluster server of the second stage of orchestration (S182) to operate a computational orchestration subroutine (S183).
[0150] The operation orchestration subroutine includes the operation orchestration stage 1, the operation orchestration stage 2, and the operation orchestration stage 3.
[0151] The first stage of computational orchestration converts application programs running on each cluster resource provider terminal connected to the distributed artificial cluster by the cluster server into containerized programs through Kubernetes processing so that they can all be used in common (S1831).
[0152] The second stage of computational orchestration schedules the allocation of resources by slump processing so that each application program can continuously use resources by the cluster server part of the first stage of computational orchestration (S1831) (S1832).
[0153] The third stage of computational orchestration processes the data transmitted and received by each application program via MPI (Message Passing Interface) by the cluster server part of the second stage of computational orchestration (S1832) to ensure consistent data communication (S1833).
[0154] If the feedback process determines that the AI data processing is not complete (S190), it feeds back to the clustering process (S160); if it determines that the AI data processing is complete (S190), it determines whether a repeating command signal is input (S200); and if it determines that a repeating command signal is not input, it proceeds to termination.
[0156] Existing artificial intelligence (AI) computations have been limited in cost and scalability due to their reliance on high-performance GPU servers or cloud resources. Meanwhile, although there are a relatively large number of PCs equipped with high-performance GPUs in companies or internet cafes, there is a lack of methods to detect these resources in real time and configure them into clusters.
[0157] High-performance GPUs in PC clients installed in PC cafes nationwide are utilized for artificial intelligence (AI) model training and inference computations. Agents are installed on the clients to collect reports on current status (whether the system is idle) as well as GPU models and specifications, which are then recorded and managed on the server in a standardized format. Based on the information managed by the server, resources required for specific tasks are analyzed and remotely clustered to be utilized for AI computation operations.
[0158] Therefore, it is possible to configure an infrastructure for AI training and computation capable of large-scale computation by utilizing the GPUs of approximately 5 million PCs nationwide without building large-scale data centers or separately securing (purchasing) GPUs for AI training.
[0159] Furthermore, by utilizing existing resources, it enables AI computation and enhances the competitiveness of national-level AI computation and learning without the need for the time and cost required to build data centers and AI servers.
[0160] Meanwhile, 1) a idle PC detection module is provided to periodically monitor PCs connected to a network and determine whether they are in an idle state by analyzing environmental information including CPU / GPU usage, login status, and power status; 2) a GPU resource metadata collection device is provided to automatically collect and standardize the GPU specifications, memory capacity, driver version, etc. of the detected idle PCs; 3) a virtual GPU cluster group is configured by considering GPU performance, network latency, and physical distance by a clustering algorithm, and is linked with computation orchestration systems such as Kubernetes, Slurm, and MPI; 4) artificial intelligence (AI) computation tasks are distributed and allocated to the GPU cluster by a resource scheduling and allocation engine, and resources are returned upon completion of the task; and 5) an authentication / encryption system is established for data protection and management of access rights to system resources by a security and authorization control module.
[0161] In other words, in terms of cost reduction, it offers the advantage of saving on the cost of introducing separate GPU servers by utilizing idle resources; in terms of flexible expansion, it secures resources in real-time based on the status of internal PCs; and in terms of eco-friendliness, it reduces power consumption and contributes to carbon emission reduction by utilizing idle equipment.
[0163] The present invention, with the above configuration, has the advantage of regularly and continuously collecting the specifications and operational information of each high-performance computer through a connected network, recording and managing them as idle computing resources, and rapidly building idle computing resources with the specifications, capacity, and performance required by a cluster processing large-scale artificial intelligence (AI) computations at the necessary time, thereby providing them at a relatively low cost.
[0165] Although the present invention has been described in detail with respect to specific embodiments described above, it is obvious to those skilled in the art that various modifications and variations are possible within the scope of the technical spirit of the invention, and it is natural that such modifications and variations fall within the scope of the appended claims. Explanation of the symbols
[0167] 90: Automatic detection of idle computing GPU resources and configured cluster system for artificial intelligence data processing 100: Cluster terminal unit 110: Terminal control unit 120: Resource Agent App Operations Department 130: Terminal Encryption Authentication Communications Department 140: Terminal frame encryption / decoding unit 150: Terminal multi-channel occupancy communication unit 160 : Terminal Touch Motion Sickness Input / Output Unit 200 : Cluster Communication Network Unit 210: Cluster Communication Network Management Department 220: Mobile Communication Channel Department 230: Mobile communication switching unit 240: Wi-Fi channel unit 250: Wi-Fi switching unit 260: Bluetooth channel unit 270: Bluetooth switching unit 280: Wired channel switching unit 290 : Cluster Network DB Section 300 : Cluster Server Section 302: Server Centralization Department 304: Server Agent App Management Department 306: Server Password Authentication Communication Section 308: Server Frame Encryption / Decryption Section 310: Idle GPU Resource Detection Unit 312: Resource Metadata Collection Unit 314: Metadata Standardization Department 316: Cluster Management Department 318 : Cluster Network Traffic Analysis Department 320 : Orchestration Management Department 322 : Server Multi-Channel Occupancy Communication Unit 400 : Cluster Resource Provision Terminal Unit 500 : Cluster Idle Resource DB 600 : Module Enclosure 610: Airflow generator 620: Exhaust port 630 : Updraft induction path 640 : Blower
Claims
Claim 1 A cluster terminal unit (100) that outputs data requiring artificial intelligence processing and corresponding request signals based on input command signals from a resource agent app that is downloaded, installed, and operated in an active state, and establishes a password authentication connection with a counterparty and performs encrypted communication; a cluster communication network unit (200) that connects to the cluster terminal unit (100), sets up switching to connect one or two communication paths with the called party based on a request signal from the calling party, and blocks the connection of the connected communication path based on a hook-off request signal from one of the connected parties; and a cluster server unit (300) that connects a portion of idle computing GPU resources detected and recorded and managed based on a request signal from the cluster terminal unit (100) connected via the cluster communication network unit (200), operates as a distributed artificial intelligence cluster for processing requested artificial intelligence data, establishes a password authentication connection with a selected counterparty and performs encrypted communication, and provides the latest version of the resource agent app. A cluster resource providing terminal (400) that connects to the cluster server unit (300) via the cluster communication network unit (200) and provides idle resource information including the specifications of GPU resources, video RAM capacity, and driver version provided by the cluster server unit (300) in response to a corresponding request signal from the cluster server unit (300); wherein the cluster server unit (300) is connected to each functional unit provided in the cluster server unit (300) and outputs a corresponding control signal to each functional unit in response to a command signal input to an operating program and a resource agent app program that are embedded and installed; and a server centralized operation unit (302). A server agent app management unit (304) that records and manages resource agent app programs by version according to the corresponding control signal of the server centralized operation unit (302), monitors that the latest version is installed and operated at the connected cluster terminal unit (100) or cluster resource provision terminal unit (400), and outputs the corresponding control signal to set the operating status by downloading and automatically installing via a streaming method; and according to the corresponding control signal applied from the server centralized operation unit (302) and the resource agent app, respectivelyA server password authentication communication unit (306) that connects to a cluster communication mode when a mutually authenticated state is confirmed by mutually exchanging connection passwords with a cluster terminal unit (100) or a cluster resource provision terminal unit (400); and a server frame encryption / decoding unit (308) that encrypts and converts a signal transmitted to the cluster terminal unit (100) or the cluster resource provision terminal unit (400) connected to the cluster communication mode by a corresponding control signal authorized by the server centralized operation unit (302) and the resource agent app, respectively, into a packet frame for transmission, and decrypts and converts a signal received as a packet frame for input. An idle GPU resource detection unit (310) that detects idle computing resources equipped with GPUs by connecting to the cluster resource provision terminal (400) in cluster communication mode by means of corresponding control signals respectively authorized from the server centralized operation unit (302) and the resource agent app, and determines whether the idle state is determined by periodically monitoring the cluster resource provision terminal (400) connected to the network and analyzing environmental information including CPU and GPU usage rates, login status, and power status; and a resource metadata collection unit (312) that collects metadata of GPU resources equipped in the idle computing resources detected by the idle GPU resource detection unit (310) by means of corresponding control signals respectively authorized from the server centralized operation unit (302) and the resource agent app; A metadata standardization unit (314) that records and manages metadata collected by the resource metadata collection unit (312) by entering it into a standardized form based on corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, respectively; and a cluster resource provision terminal unit (400) that analyzes data requesting artificial intelligence processing from the cluster terminal unit (100) based on corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, estimates the required number and specifications of GPUs, video memory capacity, model name, and driver version, and connects cluster resource provision terminal units (400) corresponding to a multiple of 1.5 in a cluster communication mode state to form and manage a distributed artificial intelligence cluster.Cluster management unit (316); a cluster network traffic analysis unit (318) that estimates the transmission error rate, transmission delay, and physical distance by transmitting and receiving test data on each communication path allocated to connect each cluster resource providing terminal (400) constituting the distributed artificial cluster formed by the cluster management unit (316) by means of corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, respectively; and an orchestration management unit (320) that organically integrates each GPU resource of each cluster resource providing terminal (400) constituting the distributed artificial cluster connected by the cluster management unit (316) by means of corresponding control signals authorized by the server centralized operation unit (302) and the resource agent app, and eliminates the occurrence of data processing bottlenecks by coordinating mutual data processing functions, and is operated by a program including Kubernetes, Slurm, and MPI (Message Passing Interface). A server multi-channel occupancy communication unit (322) that connects to the cluster communication network unit (200) by means of corresponding control signals authorized from the server centralized operation unit (302) and the resource agent app, respectively, requests the allocation of two communication paths to connect simultaneously, and transmits and receives communication signals of the same content in both directions; The above includes the server centralized operation unit (302), server agent app management unit (304), server password authentication communication unit (306), server frame encryption / decryption unit (308), idle GPU resource detection unit (310), resource metadata collection unit (312), metadata standardization unit (314), cluster management unit (316), cluster network traffic analysis unit (318), orchestration management unit (320), and server multi-channel occupancy communication unit (322), each of which is made of a printed circuit board (PCB) and is mounted in the form of a server rack on the top of the slot (SL) and the slot (SL), which is installed in a state corresponding to the inner left and right walls of the module enclosure unit (600), and is connected to the main board (B) fixed to the rear of the module enclosure unit (600) and generates heat during operation, and to air cool the heat, an airflow generator (610) is installed on both sidesEach is installed, and the airflow generator (610) is installed on the exterior of both sides of the module housing (600), and the module housing (600) is provided with a plurality of exhaust holes (620) formed by perforating them at an upward inclination on both sides, and an upward airflow inducing path (630) is formed between the airflow generator (610) and both sides of the module housing (600), and a blower (640) is installed at the bottom of the airflow generator (610) to supply air to create an upward airflow through the upward airflow inducing path (630), wherein the upward airflow inducing path (630) has a shape in which the path space widens as it goes upward, and the module housing (600) and the airflow generator (610) are spray-coated with a surface coating agent to have waterproofness, anti-contamination properties, and durability, and the surface coating agent is, per 100 parts by weight of transparent polypropylene resin, The cluster idle resource database unit (500) is composed by mixing 5 parts by weight of fine mica powder, 10 parts by weight of methylsulfonic methane, 3.5 parts by weight of sodium boroshydride, 15 parts by weight of polytetrafluoroethylene, 10 parts by weight of urea, 10 parts by weight of phosphite, 5 parts by weight of gluconate, and 10 parts by weight of sodium bicarbonate, and detects and collects idle resource information recorded in a standardized format via a corresponding control signal of the cluster server unit (300) connected via the cluster communication network unit (200), records it in an allocated memory area, and outputs it by searching; and the cluster terminal unit (100) further includes a terminal control unit (110) that connects to each functional unit provided in the cluster terminal unit (100) and outputs a corresponding control signal based on a built-in operating program, a resource agent app program, and a command signal input with data. A resource agent app operation unit (120) that connects to the terminal control unit (110), records and stores the latest version of the resource agent app program downloaded via streaming from the cluster server unit (300) by means of a corresponding control signal in an allocated memory area, and automatically installs and operates the resource agent app by means of a corresponding control signal remotely authorized from the cluster server unit (300);A terminal password authentication communication unit (130) that connects to a cluster communication mode when a mutually authenticated state is confirmed by mutually exchanging connection passwords with the cluster server unit (300) by means of corresponding control signals authorized by the terminal control unit (110) and the resource agent app, respectively; a terminal frame encryption / decoding unit (140) that encrypts and converts a signal transmitted to the cluster server unit (300) connected to the cluster communication mode by means of corresponding control signals authorized by the terminal control unit (110) and the resource agent app, respectively into a packet frame for transmission, and decrypts and converts a signal received encrypted in a packet frame for input; and a terminal multi-channel occupancy communication unit (150) that connects to the terminal control unit (110), connects to the cluster communication network unit (200) by means of a corresponding control signal, requests the allocation of two communication paths for simultaneous connection, and transmits and receives communication signals of the same content in both directions. A terminal touch input / output unit (160) that is connected to the terminal control unit (110), outputs a multimedia signal output from the cluster terminal unit (100) by means of a corresponding control signal, and inputs an external touch signal and a multimedia signal; wherein the terminal control unit (110) is configured to calculate the signal-to-noise ratio value of a signal input from the terminal multi-channel occupancy communication unit (150) to two communication paths, output a corresponding control signal to input a signal received through the best communication path, and if analyzed as having the same value, output a corresponding control signal to input a signal received through any arbitrarily selected communication path; and the cluster communication network unit (200) is configured to include a cluster communication network management unit (210) that is connected to each functional unit constituting the cluster communication network unit (200), outputs a corresponding control signal by means of an embedded installed program and operation data, and monitors. A mobile communication channel unit (220) that connects to the cluster communication network management unit (210) above, connects to the calling side and the called side respectively by means of a control signal, transmits communication signals bidirectionally using a CDMA mobile communication method, and has one or more communication paths; and to the cluster communication network management unit (210)A mobile communication switching unit (230) that connects and assigns one or more communication paths provided in the mobile communication channel unit (220) by means of a corresponding control signal to connect the calling side and the called side, and sets the switching for call connection and call blocking of the connected communication path; a Wi-Fi channel unit (240) that connects to the cluster communication network management unit (210), connects to the calling side and the called side respectively by means of a corresponding control signal, and has one or more communication paths that transmit communication signals bidirectionally using a Wi-Fi communication method; and a Wi-Fi switching unit (250) that connects to the cluster communication network management unit (210), assigns one or two communication paths provided in the Wi-Fi channel unit (240) by means of a corresponding control signal to connect the calling side and the called side, and sets the switching for call connection and call blocking of the communication path. A Bluetooth channel unit (260) connected to the cluster communication network management unit (210) and having one or more communication paths that are connected to the calling side and the called side respectively by a control signal and transmit communication signals bidirectionally using a Bluetooth communication method; a Bluetooth switching unit (270) connected to the cluster communication network management unit (210) and, by a control signal, allocates one or two communication paths provided by the Bluetooth channel unit (260) to connect the calling side and the called side and sets switching for call connection and call blocking of the communication paths; and a wired channel switching unit (280) connected to the cluster communication network management unit (210) and, by a control signal, allocates one or two wired communication paths to connect the calling side and the called side and sets switching for call connection and call blocking of the communication paths that transmit communication signals bidirectionally using a wired communication method. A cluster computing system configured for automatic detection of idle computing GPU resources for artificial intelligence data processing, comprising: a cluster network database unit (290) that connects to the cluster communication network management unit (210) and records and stores data generated during operation in the cluster communication network unit (200) and command signals input from the outside in a time series in a memory area each allocated by means of a control signal. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 delete Claim 7 delete Claim 8 A method for operating a cluster computing system comprising automatic detection of idle computing GPU resources for artificial intelligence data processing according to claim 1, wherein when it is determined that an operation command signal is input by the cluster server unit, operation data is initialized and the latest version of a resource agent app is installed on a cluster terminal unit and a cluster resource provider terminal unit connected to the cluster communication network unit; a resource detection process in which the cluster server unit of the preparation process sets a cluster communication mode, connects to the cluster resource provider terminal unit, and operates an idle GPU resource detection routine to detect idle GPU resources; a metadata collection process in which the cluster server unit of the resource detection process operates an idle resource metadata collection routine to collect corresponding GPU information from the idle GPU resources detected by the cluster server unit of the resource detection process; a standardization process in which the cluster server unit of the metadata collection process operates a metadata standardization conversion routine to record the collected information in a standardized format; and a cluster process in which the cluster management routine is operated when it is determined that a command signal for artificial intelligence data processing is input by the cluster server unit of the standardization process. A traffic process that operates a traffic environment evaluation routine by the cluster server unit of the above-mentioned cluster process; and an orchestration process (S180) that operates an orchestration routine by the cluster server unit of the above-mentioned traffic process; wherein the idle GPU resource detection routine comprises: an idle resource detection first step in which the cluster server unit connects to a cluster communication network unit and sets a cluster communication mode to detect a target of an idle GPU resource from a connected cluster resource providing terminal unit; and an idle resource detection second step in which the cluster server unit of the idle resource detection first step determines whether a target of an idle GPU resource has been detected, and determines whether it is in an idle state by analyzing environmental information including periodically monitored CPU and GPU usage rates, login status, and power status.A third idle resource detection step that selects a cluster resource providing terminal determined to have been detected by the cluster server unit of the second idle resource detection step as an idle GPU resource, verifies location information including the corresponding description information and IP address, and records and manages it in an allocated memory area; wherein the idle resource metadata collection routine comprises: a first resource metadata collection step that verifies and connects to the location information of the cluster resource providing terminal selected as an idle GPU resource in the resource detection process by the cluster server unit; a second resource metadata collection step that sets a cluster communication mode with the cluster resource providing terminal selected as an idle GPU resource by the cluster server unit of the first resource metadata collection step; and a third resource metadata collection step that collects metadata of the idle GPU resource connected in cluster communication mode by the cluster server unit of the second resource metadata collection step and records it in an allocated memory area in a state associated with the corresponding location information. The metadata standardization conversion routine includes: a first standardization conversion step in which idle GPU resource information of a cluster resource providing terminal, collected and recorded during the metadata collection process by the cluster server unit, is read and converted into character data recognizable by a large-scale language model of artificial intelligence and a natural language processing program; a second standardization conversion step in which the character data converted by the cluster server unit of the first standardization conversion step is recorded and managed in corresponding areas of a standardized format; and the cluster management routine includes: a first cluster management step in which the scale of idle GPU resources required to process the input data by the cluster server unit; and a second cluster management step in which the specifications, video RAM capacity, model, and version of the GPUs for configuring the distributed artificial intelligence cluster of the analyzed processing scale by the cluster server unit of the first cluster management step, and the required quantity of idle GPU resources consisting of cluster resource providing terminals.A cluster management step 3, wherein the cluster server unit of the cluster management step 2 selects idle GPU resources corresponding to 1.5 times the estimated required quantity and designates them as a cluster group; and the traffic environment evaluation routine comprises: a traffic environment evaluation step 1, wherein the cluster server unit requests and connects to the cluster resource providing terminal unit constituting the cluster group designated in the cluster process and the cluster communication network unit to the allocated communication path; a traffic environment evaluation step 2, wherein the cluster server unit of the traffic environment evaluation step 1 transmits test data to the connected cluster resource providing terminal unit and inputs the received signal; a traffic environment evaluation step 3, wherein the cluster server unit of the traffic environment evaluation step 2 analyzes the signal-to-noise ratio of the input signal to measure the data transmission error rate of the corresponding communication path; and a traffic environment evaluation step 4, wherein the cluster server unit of the traffic environment evaluation step 3 analyzes the delay time of the input signal to measure the connection delay time of the corresponding communication path. A traffic environment evaluation step 5 that measures the physical distance of the corresponding communication path by analyzing node information through which a signal input by the cluster server unit of the traffic environment evaluation step 4 has passed; and the orchestration routine comprises: an orchestration step 1 that analyzes values measured in the traffic process by the cluster server unit and sequentially sorts idle GPU resources in order from lowest to highest, starting with those with superior traffic environments; an orchestration step 2 that forms a distributed artificial cluster by connecting idle GPU resources corresponding to one multiple of the required quantity from the top order of the sequentially sorted idle GPU resources by the cluster server unit of the orchestration step 1 via a network and utilizes them for artificial intelligence data processing; and an orchestration step 3 that operates a computation orchestration subroutine by operating the distributed artificial cluster connected via a network by the cluster server unit of the orchestration step 2.The computation orchestration subroutine includes: a first computation orchestration step in which the cluster server unit converts application programs, each operating on a cluster resource provider terminal connected to the distributed artificial intelligence cluster, into containerized programs by processing them in Kubernetes so that they can all be used in common; a second computation orchestration step in which the cluster server unit of the first computation orchestration step schedules the allocation of resources by slump processing so that each application program can continuously use resources; a third computation orchestration step in which the cluster server unit of the second computation orchestration step processes data transmitted and received by each application program via MPI to ensure consistent data communication; and a fourth computation orchestration step in which the cluster server unit of the third computation orchestration step releases and returns the allocation of each idle GPU resource constituting the distributed artificial intelligence cluster when it is determined that the artificial intelligence data processing task is completed. A method for operating a cluster computing system configured with automatic detection of idle computing GPU resources for artificial intelligence data processing, including Claim 9 delete Claim 10 delete Claim 11 delete Claim 12 delete Claim 13 delete Claim 14 delete Claim 15 delete
Citation Information
Patent Citations
Method and system for cloud service based on mediation of processing power
KR101779712B1
System and method for design, tracking, measurement,prediction and optimization of data communicationnetworks
KR1020030059152A
Method for selecting communication path by considering transieving time for signaling , Apparatus thereof
KR1020130022764A
Method of providing a system for checking the location of distribution lines in underground distribution lines
KR102713927B1
Total solution providing system for supressing generation, detecting and preventing distribution of deepfake
KR102859247B1
