Method for quasi-optimal assignment of priorities for use of processors and input / output devices on computing systems with astra linux operating system

Abstract

FIELD: computing systems; operating systems.SUBSTANCE: invention relates to the field of optimizing the distribution of central processing unit (CPU) resources of computing systems running under the Astra Linux family of operating systems. The technical result consists in increasing the productivity and efficiency of computing systems by ensuring optimal use of processor power and input / output devices under conditions of varying load intensities caused by the execution of information and computing tasks (ICT). To achieve this, a dynamic redistribution of the used performance of the central processor or processors is performed between information-calculating tasks in order to minimize the delay in the execution of information-calculating tasks. In order to manage the allocated performance, priorities are reassigned to processes in the operating system using the nice parameter and the maximum number of read and write operations for input / output devices is limited based on the analysis of residual performance.EFFECT: increasing the productivity and efficiency.1 cl

Description

[0001] This invention relates to optimizing the allocation of central processing unit (CPU) resources in computing systems running the Astra Linux operating system. This technical solution is aimed at increasing the performance and efficiency of computing systems by ensuring optimal utilization of processor capacity and input / output devices under varying load intensities caused by the execution of information-processing tasks (IPT). To achieve the stated effect, the used performance of the central processor or processors is dynamically redistributed between information-processing tasks in order to minimize the latency of their execution. At the same time, to manage the allocated performance, priorities are reassigned to processes in the operating system and the maximum number of read and write operations for I / O devices is limited.

[0002] The proposed invention solves the pressing practical problem of optimizing computing resource utilization in secure computing systems based on the Astra Linux operating system. This issue is relevant due to the increasing demands on the performance, reliability, and predictability of computing systems under strict constraints on the latency of critical operations.

[0003] Computer systems with the Astra Linux operating system are often used for:

[0004] critical data processing (expert systems, decision support systems);

[0005] Monitoring and management of critical infrastructure facilities;

[0006] cryptographic operations with strict execution time requirements;

[0007] processing streaming data with a fixed frequency of receipt (e.g. sensor data, video streams).

[0008] For all these applications, the key requirement is guaranteed minimal execution latency of critical threads, regardless of the background load on the system.

[0009] The proposed invention provides a mathematically sound resource allocation to minimize latency. This allows developers of mission-critical applications to provide response time guarantees, which is impossible with standard nice mechanisms, which provide only relative priorities rather than absolute guarantees.

[0010] In multiprocessor computing systems with intensive I / O load, the “uncertainty delay” effect occurs:

[0011] The critical thread latency is unpredictable because it depends on the current state of the disk subsystem and the priorities of other threads;

[0012] The standard tools (top, htop, nice) do not provide a mechanism for coordinated control of CPU and I / O;

[0013] The administrator is forced to manually reconfigure priorities when the load changes, which requires constant monitoring and intervention.

[0014] The proposed invention automates this procedure by calculating optimal resource utilization rates based on residual performance. This is achieved without administrator intervention and allows the system to automatically adapt to changing loads, ensuring stable latency.

[0015] Modern computing systems are often built on the basis of:

[0016] multi-core processors (Intel Xeon, AMD EPYC) with the number of cores from 32 to 128;

[0017] NUMA (Non-Uniform Memory Access) systems with multiple processor sockets and local memory;

[0018] hybrid architectures with different types of cores (P-core, E-core in Intel);

[0019] high-speed I / O devices (NVMe SSD, InfiniBand, 40G+Ethernet).

[0020] On such systems, traditional prioritization mechanisms (nice) become ineffective:

[0021] Priority applies to the core where the process is running, but not to other cores;

[0022] In a NUMA system, accessing memory from another socket can be 2-3 times slower;

[0023] I / O operations compete for limited PCIe bus and memory bandwidth.

[0024] The proposed invention takes into account the residual performance of all types of resources (CPU, local memory, remote memory, I / O devices) and calculates coefficients that minimize latency in this complex multi-resource environment. This is critical for maximizing the efficiency of expensive computing systems.

[0025] The main idea of ​​the invention is to read information about the processor load, read from the contents of the system files of the Astra Linux operating system of the computing complex information about the number of read and write operations performed on all input / output devices for all optimized flows and then determine, using the residual performance of the processor and input / output devices, such resource utilization factors that will ensure a minimum of delay, and the obtained factors are used in establishing priorities for the use of the processor and the maximum intensities of reading and writing operations of the optimized flows with input / output devices.

[0026] The prior art discloses a “Method for determining the priority of tasks in a server system queue” (RU patent No. 2638635, 2014) [1].

[0027] The essence of this analogous method is as follows.

[0028] In real time, determine the priority of multiple tasks from multiple users by analyzing the cumulative CPU time, task volume, and time intervals between queues; assign a numerical priority to each user; based on the data of previous tasks and current loads, calculate metrics such as CPU time, time interval, and task volume over a specified period; adjust the user priority based on the calculated values; assign an updated priority to each task; tasks with higher priority are selected for processing first; the processor processes tasks sequentially, eliminating parallel execution.

[0029] The analog method has the following disadvantages.

[0030] 1. They use FIFO queues to process data, but there is no mechanism for managing thread priorities at the operating system level (for example, setting nice values ​​or limiting the intensity of I / O operations).

[0031] 2. Do not monitor current CPU load, read I / O information, or calculate residual device performance.

[0032] 3. They optimize data processing based on quality and processing order criteria, but do not address the issue of minimizing delays in the execution of information and calculation tasks in the operating system.

[0033] The prior art includes a “Method for dynamic quasi-optimal distribution of performance when processing real-time fractal traffic in systems with failures” (RU patent No. 2742038, 2020) [2].

[0034] This technical solution, although aimed at distributing resources when transmitting traffic, upon closer examination distributes the clock frequency of the router processor when processing network packets with data, which can be generally represented as computing tasks in the router's operating system.

[0035] The essence of this analogous method is as follows:

[0036] define the characteristics of real-time fractal traffic under no-buffering conditions;

[0037] Evaluate the long-term dependence and pulsation of the input flow using fractal indicators;

[0038] quasi-optimally distribute the available equipment performance, taking into account the current statistical parameters of traffic;

[0039] implement statistical multiplexing of packets in time intervals between adaptations of the processing system parameters;

[0040] Smooth out sudden bursts of single-priority real-time traffic in systems with failures, based on the analysis of the correlation properties of the input flow.

[0041] The analog method has the following disadvantages.

[0042] 1. Optimization is performed at the network level, not at the operating system level.

[0043] 2. Manage data transfer over network channels (Ethernet, IP), but do not control read and write operations with local storage devices (hard drives, solid-state drives).

[0044] 3. They do not read operating system files to obtain information about the state of computing resources, as they operate at the network protocol level.

[0045] 4. Manages network packet transmission queues, but does not set CPU usage priorities (nice values) for operating system threads.

[0046] 5. Optimizes latency and throughput for transmitting network packets in satellite and terrestrial networks, but does not minimize the latency of computing threads within the operating system.

[0047] A close analogue in technical essence to the claimed method is the hardware-implemented “Method for task scheduling (variants), task scheduling system and machine-readable medium” (RU patent No. 2453901, 2007) [3].

[0048] The essence of this analogous method is as follows: determine the load of the i-th computing resource; compare the obtained load levels with the established threshold values; determine which tasks can be scheduled immediately and which should be postponed; calculate resource priorities based on their relative speed; check the need for access to additional resources; schedule the execution of computing tasks taking into account the corresponding threshold criteria; record information on the actual load for analysis during subsequent planning.

[0049] The analog method has the following disadvantages.

[0050] 1. They exclusively use processor priority management, balancing priorities in the range of 20-120% of the base value. The description does not include a mechanism for monitoring or controlling read and write operations to I / O devices (hard drives, SSDs, network interfaces).

[0051] 2. They adjust priorities based on processor load and do not analyze the remaining performance of I / O devices. This leads to situations where a high-priority processor thread is blocked by a slow I / O device, increasing overall system latency.

[0052] 3. Use external APIs to initiate thread priority changes. This means that an external application or module must periodically poll the system and call the API to make adjustments, which introduces additional overhead and latency.

[0053] 4. Use heuristic rules for balancing priorities (e.g., increasing by 20% or 80-90% of the range) and do not formulate a mathematical optimization problem with an objective function of minimizing delay.

[0054] 5. Does not take into account the time required to switch the CPU between different tasks.

[0055] 6. They do not take into account the specifics of the system files, scheduling mechanisms, and prioritization of the Astra Linux operating system, which limits the possibility of its effective use in secure computing systems based on Astra Linux.

[0056] The closest in technical essence to the claimed method and chosen as a prototype is the “Method for optimally assigning priorities to processes in the Astra Linux operating system on computing complexes” (RU patent No. 2852520, 2025) [4].

[0057] The essence of this prototype method is as follows:

[0058] connect a server to a computing complex and provide remote control and network file exchange between them, using the SSH protocol for remote control and the SMB protocol for network file exchange;

[0059] run a resource usage control utility on the computing complex by sending a command from the server, using the top or htop utilities as the resource usage control utility;

[0060] saves the content displayed on the screen by the utility for monitoring the use of computing complex resources to a file on the server every second;

[0061] read the contents of the file on the server in real time and programmatically determine the percentage of residual performance of the central processor after executing the current system processes, while the current workload of the computing complex is read from the file and the idle clock frequency is calculated;

[0062] collect statistics on the server in real time about the burstiness of the volume of required computing resources for all processes involved in the redistribution of the clock frequency, through continuous software analysis of the contents of a file with information about the system load per second, while burstiness is understood, as is accepted in the theory of deterministic network calculus, as the maximum surge in the cumulative function for the volume of the processor clock frequency required by the process;

[0063] Determine on the server, by a complete enumeration method, the coefficients of utilization of the residual computing power that will ensure the minimum sum of the ratio of the process burstiness to the product of the coefficient of utilization of the residual clock frequency and the total residual clock frequency of the central processor summed with the switching time between processes;

[0064] Calculates priorities on the server in the Astra Linux operating system by normalizing the coefficients of utilization of the residual clock frequency of the central processor found by the exhaustive search method;

[0065] change the value of the “nice” parameter in the computing complex for the optimized processes in the Astra Linux operating system by sending a command from the server, using the “renice” command transmitted via the SSH protocol;

[0066] Recalculate and reassign the value of the "nice" parameter when processes appear or end;

[0067] assign the highest priority to the process of each new information calculation task.

[0068] The prototype method has the following disadvantages.

[0069] 1. Manage processor priorities using the nice mechanism in the Astra Linux operating system (the nice value range is from -20 to +19), using the top, htop, and renice tools without monitoring or limiting the intensity of read and write operations with I / O devices.

[0070] 2. Analyze processor load using the top and htop utilities and do not read information about the number of read and write operations performed with input / output devices from the operating system files.

[0071] 3. Manage CPU priorities exclusively, without integrating CPU and I / O management into a single optimization system with a common latency minimization objective.

[0072] The technical result of the proposed method consists in providing a comprehensive automatic optimization of the use of processor resources and input / output devices.

[0073] This is achieved through dynamic optimal redistribution of the used CPU performance and the performance of input / output devices between information-calculating tasks with mathematical calculation of coefficients to minimize latency based on reading system files of the Astra Linux operating system.

[0074] The problem solved by the proposed method consists of assigning priorities to information and computing tasks performed on computing systems running the Astra Linux operating system in an optimal manner, based on the criterion of minimum total delay, taking into account the residual performance of the processor or processors and the residual performance of the input / output devices used.

[0075] To solve the stated problem, a method for quasi-optimal assignment of priorities to processes for using the processor and input / output devices on computing complexes with the Astra Linux operating system is proposed.

[0076] The essence of the proposed method is as follows.

[0077] 1. Connect a priority management server to the computing complex and provide remote control and network file sharing between them.

[0078] In this case, the SSH protocol is used for remote control, and the SMB or FTP protocol is used for network file sharing.

[0079] 2. Launch the resource usage control utility on the computing complex by sending a command from the priority management server.

[0080] In this case, top or htop utilities are used as a resource usage control utility.

[0081] 3. Save the content displayed on the screen by the computing complex resource usage monitoring utility to a file on the priority management server every second.

[0082] This uses the flow redirection operation for the top or htop command.

[0083] 4. Read the file contents on the server in real time and programmatically determine the percentage of residual CPU performance after executing current system processes.

[0084] The current workload of the computing complex is read from the file and the idle clock frequency is calculated. To calculate the residual performance as a clock frequency, use an expression like this:

[0085] where - CPU clock frequency (total clock frequency of all processors); PR CPU - the percentage of the processor or processors occupied by system or utility programs.

[0086] 5. Transfer copies of the Astra Linux operating system files containing information about read and write operations performed with all input / output devices in real time, every second, from the computing complex to the priority management server via the SMB or FTP protocols.

[0087] 6. Read the number of read and write operations performed in one second for all I / O devices from the received system files of the Astra Linux operating system on the priority management server and programmatically determine the residual performance separately for read and write operations.

[0088] In this case, expressions of the following type are used:

[0089]

[0090]

[0091] where λ j (max_r) - the maximum possible reading intensity from the j-th input / output device, λ, based on the tactical and technical characteristics j(max_w) - the maximum possible recording intensity from the tactical and technical characteristics on the j-th input / output device, λ j (r) - current reading intensity from the j-th input / output device, λ j (w) - current write intensity on the j input / output device.

[0092] 7. The priority management server calculates the intensity of read and write operations performed by all optimized information and computing tasks with all input / output devices by analyzing system files transferred from the computing complex.

[0093] 8. Collect statistics on the burstiness of the required computing resources for all processes participating in the distribution of the residual clock frequency in real time on the priority management server through continuous software analysis of the contents of a file with information on the system's per-second workload.

[0094] Burstiness, as is commonly understood in deterministic network computation theory, is the maximum burst of the cumulative function for the amount of processor clock frequency required by a process. Using burstiness allows for accounting for jitter in the CPU frequency required for the IRP.

[0095] 9. Determine on the priority management server, using an exhaustive search method, the residual performance utilization factors and the I / O device utilization factors for reading and writing that will ensure the minimum sum of the ratio of process burstiness to the product of the residual clock frequency utilization factor and the total residual clock frequency of the central processor, summed with the process switching time and summed with the latency of reading and writing operations for all optimized threads and the I / O devices they use.

[0096] For this, an expression of the following type is used:

[0097]

[0098] where b i - the burstiness of the required volume of calculations for the i-th information-calculation task, - switching time between processes, proc - the number of processes of information-calculation tasks participating in optimization, 10 - the number of input / output devices with which the optimized information-calculation tasks work, λ j (ост_r) - residual reading intensity from the j-th input / output device, λ j (ост_w) - residual recording intensity on the j-th input / output device, λ ij (r) - current reading intensity of the i-th IRZ from the j-th input / output device, λ ij (w) - current intensity of writing the i-th IRZ to the j-th input / output device, T j (r) - the execution time of one read operation from the j-th input / output device, T j (w) - the execution time of one write operation to the j-th input / output device.

[0099] In this case, the following expression is used as the objective function for minimization:

[0100]

[0101] 10. Calculate processor usage priorities for the Astra Linux operating system on the computing complex for optimized information and computational tasks on the priority management server by normalizing the residual clock frequency utilization factors of the central processor found by the exhaustive search method.

[0102] In this case, for normalization, that is, for converting the value of the residual clock frequency utilization coefficients, which take values ​​from 0 to 1, to the value of the nice priority parameter, an expression of the following type is used:

[0103]

[0104] where - the coefficient of utilization of the residual computing power for the i-th process in the operating system of the computing complex, which is involved in the optimization of computing resources.

[0105] 11. Change the "neck" parameter value for optimized processes in the Astra Linux operating system by sending a command from the priority management server. This is done using the "renice" command transmitted via the SSH protocol.

[0106] 12. Recalculate and reassign the value of the "nice" parameter when processes appear or end.

[0107] 13. Change the maximum permitted intensity of read and write operations for each input / output device in the computing complex for all optimized information and computing tasks.

[0108] 14. Assign the highest priority to the process of each new information calculation task.

[0109] 15. Assign the maximum possible read and write performance to all input / output devices for each new information processing task.

[0110] The industrial applicability of the method is due to the possibility of implementing it on any computing complexes running the Astra Linux operating system, as well as on any computing complex capable of dynamically redistributing the performance of the central processor or processors, as well as the resource of input / output devices between computing tasks.

[0111] The following main actions are performed on the material object "computing complex" and "priority management server": connecting the priority management server to the computing complex with the ability to be controlled remotely, running the resource usage control utility, saving the CPU usage statistics to a file, reading the CPU usage statistics from a file, sending a remote command to change the priority of processes, transferring the contents of system files, and changing the contents of system files.

[0112] A comparison of the claimed Method for quasi-optimal assignment of priorities to processes for using the processor and input / output devices on computing complexes with the Astra Linux operating system and the prototype shows that the claimed method differs significantly from the prototype.

[0113] General features of the claimed method and prototype:

[0114] 1. Connect a priority management server to the computing complex and provide remote control and network file sharing between them.

[0115] 2. Launch the resource usage control utility on the computing complex by sending a command from the priority management server.

[0116] 3. Save the content displayed on the screen by the computing complex resource usage monitoring utility to a file on the priority management server every second.

[0117] 4. Read the file contents on the server in real time and programmatically determine the percentage of residual CPU performance after executing current system processes.

[0118] 5. Collect statistics on the burstiness of the required computing resources for all processes participating in the distribution of the residual clock frequency in real time on the priority management server by continuously analyzing the contents of a file with information on the system's per-second workload.

[0119] 6. Determine the coefficients of utilization of the residual performance of the central processor on the priority management server using the exhaustive search method.

[0120] 7. Calculate processor usage priorities for the Astra Linux operating system on the computing complex for optimized information and computational tasks on the priority management server by normalizing the residual clock frequency utilization factors of the central processor found by the exhaustive search method.

[0121] 8. Change the value of the "nice" parameter for optimized processes in the Astra Linux operating system in the computing complex by sending a command from the priority management server.

[0122] 9. Recalculate and reassign the value of the "nice" parameter when processes appear or end.

[0123] 10. Assign the highest priority to the process of each new information calculation task.

[0124] Distinctive features of the proposed solution:

[0125] 1. Copies of the Astra Linux operating system files containing information on read and write operations performed with all input / output devices are transmitted in real time every second from the computing complex to the priority management server via SMB or FTP protocols.

[0126] 2. Read the number of read and write operations performed in one second for all I / O devices from the received system files of the Astra Linux operating system on the priority management server and programmatically determine the residual performance separately for read and write operations.

[0127] 3. Calculate on the priority management server, by analyzing system files transferred from the computing complex, the intensity of read and write operations performed by all optimized information and computing tasks with all input / output devices.

[0128] 4. Determine the read and write utilization rates of I / O devices on the priority management server using the exhaustive search method.

[0129] 5. Change the maximum permitted read and write intensity for each input / output device in the computing complex for all optimized information and computing tasks. This includes making changes to system files using the SSH protocol.

[0130] 6. Assign the maximum possible read and write performance to all input / output devices for each new information processing task.

[0131] Thus, the proposed invention is specifically adapted for Astra Linux and utilizes the OS's system mechanisms (reading / proc files, managing nice priorities, and integrating with the Linux kernel). This enhances the ability to provide high-performance computing with guaranteed latency characteristics.

[0132] Astra Linux, as a Linux-based OS, provides access to detailed information about the state of computing resources through system files in the / proc directory:

[0133] / proc / [pid] / io - information about read and write operations for each process (read_bytes, write_bytes);

[0134] / prcos / stat - CPU load statistics by cores;

[0135] / proc / diskstats - information about I / O operations for each block device;

[0136] / sys / block / [device] / stat - detailed performance metrics of storage devices.

[0137] The proposed invention effectively utilizes this unique feature of Astra Linux systems by reading actual I / O data in real time. This allows optimization to be based on accurate, factual data rather than heuristic estimates. No proprietary operating system (Windows, macOS) provides such complete access to performance metrics at the user process level.

[0138] In the Linux kernel (and, accordingly, in Astra Linux), thread prioritization occurs through the nice mechanism, which determines the share of CPU time allocated to each process. Nice values ​​range from -20 (maximum priority) to +19 (minimum priority).

[0139] However, managing only CPU priorities without I / O control leads to the following problems:

[0140] High-priority thread blocking on I / O: A thread with priority -20 may be blocked on reading from a slow disk, which will negate the effect of high priority;

[0141] I / O device contention: Multiple threads with different priorities compete for access to the same disk, and the kernel may schedule operations suboptimally;

[0142] No dynamic balancing possible: priorities are set statically, but I / O load changes dynamically.

[0143] The proposed invention solves this problem through comprehensive optimization: based on an analysis of the current processor load and the remaining performance of I / O devices, both CPU priorities (nice) and maximum I / O operation intensities (via cgroups, blkio, and other mechanisms) are automatically adjusted. This ensures predictable execution latency for critical threads.

[0144] SOURCES OF INFORMATION

[0145] 1.Pat. 2453901 Russian Federation, IPC G06F 9 / 46. Hardware-implemented method for task scheduling (variants), task scheduling system and machine-readable medium / Jensen K., Steffer A., ​​Thomas B., Kedruvey R.; patent holder Discier Corporation (US). - No. 2008149050 / 08; declared 06.06.2007; published 20.06.2012, Bulletin No. 17. - 18 p.

[0146] 2. Patent. 2638635 Russian Federation, IPC G06F 9 / 48. Method for determining the priority of tasks in the queue of a server system / Panferov V.V., Meshcheryakov D.K.; patent holder OOO "Abi Development" (RU). - No. 2014125149; declared. 20.06.2014; published. 14.12.2017, Bulletin No. 35.-18 p.

[0147] 3. Patent. 2742038 Russian Federation, IPC H04L 47 / 10. Method for dynamic quasi-optimal distribution of performance when processing real-time fractal traffic in systems with failures / Martyanov A. N., Belov P. Yu., Voronin O. I.; patent holder Peter the Great Strategic Missile Forces Military Academy (RU). - No. 2020111825; declared 23.03.2020; published 02.02.2021, Bulletin No. 4. - 7 p.

[0148] 4. Patent. 2852520 Russian Federation, IPC G06F 9 / 505, G06F 12 / 02. Method for optimal assignment of priorities to processes in the ASTRA LINUX operating system on computing complexes / Sytsevich N. F., Belov P. Yu., Khomchenkov A. V.; patent holder Peter the Great Strategic Missile Forces Military Academy (RU). - No. 2025116596; declared 17.06.2025; published 09.12.2025, Bulletin No. 34. - 6 p.

Claims

A method for quasi-optimal assignment of priorities to processes for using the processor and input / output devices on computing systems with the Astra Linux operating system, which consists of: connect a priority management server to the computing complex and provide the ability to remotely control and network file exchange between them; launch a resource usage control utility on the computing complex by sending a command from the priority management server; save, at intervals of one second, the content displayed on the screen by the utility for monitoring the use of computing complex resources to a file on the priority management server; read the file contents on the server in real time and programmatically determine the percentage of residual CPU performance after executing current system processes; collect statistics on the priority management server in real time about the burstiness of the volume of required computing resources for all processes involved in the distribution of the residual clock frequency, through continuous software analysis of the contents of a file with information on the system's per-second workload; determine the coefficients of utilization of the residual performance of the central processor on the priority management server using the exhaustive search method; On the priority management server, the calculation of the priorities for the use of the processor for the Astra Linux operating system on the computing complex is carried out by optimizing the information and computational tasks by normalizing the coefficients of use of the residual clock frequency of the central processor found by the exhaustive search method; change the value of the “nice” parameter in the computing complex for the processes being optimized in the Astra Linux operating system by sending a command from the priority management server; recalculate and reassign the value of the "nice" parameter when processes appear or end; assign the highest priority to the process of each new information-calculation task, characterized by the fact that: transmit in real time with a frequency of one second to the priority management server from the computing complex via SMB or FTP protocols copies of the system files of the Astra Linux operating system, containing information about the read and write operations performed with all input / output devices; read from the received system files of the Astra Linux operating system on the priority management server the number of read and write operations performed in one second for all input / output devices, and programmatically determine the residual performance separately for read and write operations; on the priority management server, by analyzing data from the computing complex of system files, the intensity of read and write operations performed by all optimized information and computing tasks with all input / output devices is calculated; determine the coefficients of usage of input / output devices for reading and writing on the priority management server using the exhaustive search method; change the value of the maximum permitted intensity of read and write operations for each input / output device in the computing complex for all optimized information and computing tasks; assign to the process of each new information-calculation task the maximum possible intensities of work on reading and writing with all input / output devices.

Images