2687-0517
Computing, Telecommunication and Control
17
3
2024
1-152
RAR
RUS
9-21
Peter the Great St.Petersburg Polytechnic University
Vladimir
S.
vlad@neva.ru
Polytechnicheskaya, 29, St.Petersburg, 195251, Russia
F-6480-2013
7004013271
0000-0002-5637-1420
Peter the Great St.Petersburg Polytechnic University
Lev
V.
lev.utkin@mail.ru
Polytechnicheskaya, 29, St.Petersburg, Russia, 195251
Muliukha
V.A.
mva@rtc.ru
Exo-intelligent hybrid supercomputer platforms for shared-use centers
The article discusses the possibilities of increasing the real performance of hybrid supercomputer platforms consisting of different types of processor nodes (CPU, GPU, FPGA) operating in the mode of shared-use computational resources. The conceptual difference of the proposed approach from widespread supercomputing cluster platforms can be metaphorically expressed as “Less Moore, more brain.” The considered approach shifts the focus of technology development from classical methods of increasing the performance of HPC platforms by adding new hardware multi-core computing components to more complex exo-intelligent solutions that use inductive (internal) and conceptual (external) data to implement machine learning methods for the purpose of optimally distributing available hardware resources between different classes of user applications. The proposed three-level architecture of hybrid computing platforms opens up new opportunities both for efficient scaling of user program execution processes, and for reification of descriptions of new algorithms by generating corresponding texts of computer programs, as well as interpreting the results obtained based on the use of statistical information, the carrier of which is censored data characterizing the experience of executing user applications in the mode of shared use of hybrid computational resources.
10.18721/JCSTCS.17301
004
high performance hybrid computing systems
machine learning
scheduler
survival function
explainable artificial intelligence
https://infocom.spbstu.ru/article/2024.82.1/
RAR
RUS
22-31
F-6480-2013
7004013271
0000-0002-5637-1420
Peter the Great St.Petersburg Polytechnic University
Lev
V.
lev.utkin@mail.ru
Polytechnicheskaya, 29, St.Petersburg, Russia, 195251
Peter the Great St. Petersburg Polytechnic University
Konstantinov
Andrei
andrue.konst@gmail.com
St. Petersburg, Russian Federation
0000-0002-1115-7543
Peter the Great St. Petersburg Polytechnic University
Eremenko
Danila
eremenko.dyu@edu.spbstu.ru
St. Petersburg, Russian Federation
Peter the Great St.Petersburg Polytechnic University
Vladimir
S.
vlad@neva.ru
Polytechnicheskaya, 29, St.Petersburg, 195251, Russia
Muliukha
V.A.
mva@rtc.ru
Interpretation methods for machine learning models in the framework of survival analysis with censored data: a brief overview
Methods of interpretation, or explanation, of predictions are an integral part of modern black-box machine learning models. They have become widespread due to the need for the user to understand what the machine learning model is predicting. This is especially important for survival analysis models, as they are used in medicine, system reliability, safety, and also have features that make them difficult to explain and interpret. The paper discusses the main methods for interpreting survival models that deal with censored data and determine the characteristics of the time until a certain event. A feature of such models is that their predictions are presented not as a point value, but as a probabilistic function of time, for example, a survival function or a risk function. This requires the development of special interpretation methods. The most well-known methods SurvLIME, SurvLIME-KS, SurvNAM and SurvBeX, SurvSHAP(t) are considered, which are based on the use of LIME and SHAP interpretation methods, the Cox model and its modifications, as well as the Beran estimator.
10.18721/JCSTCS.17302
004.85
machine learning
survival model
explainable artificial intelligence
censored data
Cox model
Beran estimator
https://infocom.spbstu.ru/article/2024.82.2/
RAR
RUS
32-41
Muliukha
V.A.
mva@rtc.ru
Vostrov
Alexey
alex.sinkriver@gmail.com
Motorin
Dmitrii
d.e.motorin@gmail.com
Glazunov
Vadim
neweagle@gmail.com
Supercomputer resources management using machine learning methods under constraints
Artificial intelligence and machine learning technologies are among the most promising in the field of computer science. They make it possible to obtain solutions to problems that until recently were the exclusive prerogative of humans. However, when solving practical problems, it is necessary to implement machine learning models taking into account the restrictions on available resources. Such resources can be both computational and temporary (i.e. the problem must be solved in a certain time and using certain hardware, most often it is about various mobile platforms), and informational, when it comes to small, censored, incomplete or noisy data. The paper examines machine learning methods used to solve practical problems in application areas, such as comparing the shape of three-dimensional objects and intellectualizing resource dispatching, within the framework of the concept of “Supercomputer for AI and AI for a Supercomputer”. In the field of solving problems with limited data volume, a method is proposed that allows training a multilayer neural network using an ultra-small training sample to solve the problem of quantitatively assessing the proximity of the shape of arbitrary three-dimensional objects. In the field of applying machine learning models with limited resources, a method has been developed that ensures asynchronous operation of the machine learning model and the executable process, which allows for the effective use of machine learning methods under constraints.
10.18721/JCSTCS.17303
004.8:[62-5+004.382.2]
machine learning
resource management
supercomputer
constraints
neural networks
https://infocom.spbstu.ru/article/2024.82.3/
RAR
RUS
42-53
0000-0003-0093-6506
Peter the Great St. Petersburg Polytechnic University
Malov
Sergey
St. Petersburg, Russian Federation
Lukashin
Aleksey
lukash@neva.ru
Count time series analysis of jobs scheduling in the hybrid supercomputer center
Increasing the efficiency of supercomputer centers is an extremely important task, especially in the context of growing demand for high-performance computing and a shortage of supercomputer resources. Statistical analysis of the results of various indicators of supercomputer performance is aimed at creating models of computing resource management and forming a basis for using artificial intelligence methods. The purpose of this research is to study the incoming flow of user requests (jobs), which largely determines the load on supercomputer resources. To analyze the incoming flow of user jobs, generalized linear models and generalized estimating equations, as well as the autoregressive conditional Poisson model, were used. It allowed taking
into account the dependence of observations and the effect of overdispersion. Based on the results of supercomputer operation observations, estimates of the time trend were obtained, as well as indicators of changes in the intensity of the job flow within weekly and annual cycles with classification by areas of expertise and computing clusters. Indicators of statistical significance of changes within the weekly and annual cycles were established. As a result of an advanced statistical analysis using multiple comparison methods, statistically significant orders of the main effects of the weekly and annual factors were obtained.
10.18721/JCSTCS.17304
519.25
count time series
generalized estimating equations
autoregressive conditional Poisson model
multiple comparisons
supercomputer cluster
job scheduling
https://infocom.spbstu.ru/article/2024.82.4/
SCO
RUS
54-60
Peter the Great St. Petersburg Polytechnic University
Konstantinov
Andrei
andrue.konst@gmail.com
St. Petersburg, Russian Federation
Predictive models and dynamics of estimates of applied tasks characteristics using machine learning methods
The paper considers the machine learning problem of simultaneous estimation of the conditional survival distribution and dynamic characteristics of computational tasks. The problem arises in cluster workload management and is extremely relevant for optimal scheduling. To solve the problem, a new method is proposed, based on the combination of the attention mechanism and the random survival forest. The key feature is the use of a tree structure derived from a random survival forest. The forest construction algorithm uses only the survival dataset. Each leaf uses the unconditional Kaplan-Meier estimate, which is a serious limitation of the forest, especially for rare events in some parts of the feature space. Moreover, the random survival forest does not allow estimating the dynamic parameters of the task. The proposed method solves these problems by extending the already constructed random survival forest with the attention mechanism inside each leaf of the tree. The Beran estimator is used to model survival distribution, and the Nadaraya-Watson regression with the same parameters is used to predict the dynamic characteristics of tasks. To do this, subsets of training data corresponding to the same leaf as the input vector are used. As a result, the joint model is obtained that allows us to estimate the survival function more accurately and at the same time to predict the dynamic characteristics of the task. The developed model combines the advantages of smooth models based on the attention mechanism and stepwise decision trees.
10.18721/JCSTCS.17305
004.855.5
machine learning
survival analysis
attention mechanism
random survival forest
Beran estimator
https://infocom.spbstu.ru/article/2024.82.5/
RAR
RUS
61-70
Peter the Great St. Petersburg Polytechnic University
Mezheneva
Irina
Mezheneva-I@gaz-is.ru
St. Petersburg, Russian Federation
Lukashin
Aleksey
lukash@neva.ru
Peter the Great St. Petersburg Polytechnic University
Chatoyan
Sergey
Chatoyan-S@gaz-is.ru
St. Petersburg, Russian Federation
Reconstruction of attractors of super-computer user's activity and identification of critical deviations in their behavior
The modern job scheduling system in supercomputer platforms is based on the estimates of the request for computing resources provided by users (often based on subjective considerations). However, it has been found that such estimates can be significantly inaccurate. In this regard, a practically important task arises: building a behavior model of user tasks executed in a supercomputer, identifying and evaluating critical deviations from the predicted behavior profile (based on an assessment of user confidence). Methods of nonlinear dynamics and topological data analysis are used to solve this problem. The article presents the results of experimental studies for various data sets obtained at the “Polytechnic Supercomputer Center” of Peter the Great St. Petersburg Polytechnic University. The Betti curves of the supercomputer user profile are calculated. The results of the evaluation of the comparison of several user profiles with the reference profile are presented. A desirability scale and numerical intervals for the proposed classes are proposed.
10.18721/JCSTCS.17306
004.8
high performance systems
hybrid computing systems
topological data analysis
scalar time series
job scheduling
https://infocom.spbstu.ru/article/2024.82.6/
RAR
RUS
71-83
Zayats
Oleg
zay.oleg@gmail.com
Peter the Great St. Petersburg Polytechnic University
Baksheev
Vitaliy
St. Petersburg, Russian Federation
Peter the Great St.Petersburg Polytechnic University
Vladimir
S.
vlad@neva.ru
Polytechnicheskaya, 29, St.Petersburg, 195251, Russia
Muliukha
V.A.
mva@rtc.ru
Model of a supercomputer cluster in the form of a queueing system with a random limit on the execution time of applied tasks
It is well known that the efficiency of task dispatching in any supercomputer system is determined, first of all, by the adequacy of the system model used, as well as the accuracy of the estimation of the parameters of the model itself. The article proposes a new version of the supercomputer cluster model, based on a standard model of the M/M/∞ class queueing system, which is supplemented with two fundamental clarifications that reflect the features of the supercomputer operation. First, the processing time of each task is limited by the dispatcher using a random variable distributed according to the exponential law. Second, it is considered that each new task requires the allocation of a random number of service channels (processors) for its execution. The parameters of the proposed queueing model are estimated based on statistical processing of data obtained during calculations previously performed on a supercomputer. A number of examples of using the developed model are given. To calculate the parameters of the queueing system, it is proposed to use the method of generating functions.
10.18721/JCSTCS.17307
519.872.4:004.451.44
queueing system
task dispatching
supercomputer
random execution time
random number of service channels
https://infocom.spbstu.ru/article/2024.82.7/
RAR
RUS
84-92
Peter the Great St. Petersburg Polytechnic University
Antonenko
Maksim
St. Petersburg, Russian Federation
Budanov
Dmitriy
dmitriy.budanov@gmail.com
0000-0003-0334-2770
Institute for Analytical Instrumentation of the Russian Academy of Sciences
Zaitceva
Anna
St. Petersburg, Russian Federation
Non-invasive heart rate measurement system based on video stream analysis
The paper is devoted to the development and testing of a remote biomonitoring system based on the phenomenon of plethysmography. This phenomenon allows not only to measure a person’s pulse rate non-invasively, but also to assess physiological state of the person. At the first stage of the system operation, it is necessary to detect regions of interest. This operation can be effectively implemented using neural networks. The task of face recognition was performed by the YOLOv7-tiny architecture, due to its speed and the ability to run on embedded systems. For the detected face, a rectangle was created, whose coordinates indicated the boundaries of the face. Next, the average brightness of the selected areas is calculated and stored in the dataset. By performing fast Fourier transform (FFT) for a given set, it is possible to obtain a signal spectrum. Using methods of digital signal processing, it is possible to filter the signal and select the part of the spectrum of interest in the region of 0.7–3 Hz. The maximum amplitude of the harmonic will correspond to the current pulse.
10.18721/JCSTCS.17308
004.89
remote biomonitoring
telehealth
heart rate
photoplethysmography
Fourier transform
neural network
https://infocom.spbstu.ru/article/2024.82.8/
RAR
RUS
93-102
Peter the Great St. Petersburg Polytechnic University
Lapin
Igor
lapin_ia@spbstu.ru
St. Petersburg, Russian Federation
Peter the Great St.Petersburg Polytechnic University
Sabinin
Oleg
olegsabinin@mail.ru
Polytechnicheskaya, 29, St.Petersburg, 195251, Russian Federation
Development of the system of automatic generation of database model on the basis of the task text in natural language
This paper describes an approach to the implementation of a system that would allow automatic database model generation from a natural language description given by the user. Different machine learning technique, such as transformer, named entity recognition and relation extraction are considered and applied. The implementation of the neural network model uses the capabilities of the spaCy framework to organize a generic pipeline for training. Off-the-shelf implementations of some individual components from spaCy are also used, while the rest are custom. Moreover, we describe the process of gathering and preparing raw data for training a neural network model, and generating a proper corpus from them. For this purpose, a specialized annotating tool, Doccano, is used, which satisfies all requirements and is freely available. Finally, the paper presents the model parameters used in training and the performance metrics obtained. We’ve been able to achieve great results for the named entity recognition component, while the performance metrics of the relation extraction component can still be improved. The paper concludes with possible directions for further work on the implementation of the described system, including the relation extraction component improvements and new features implementation.
10.18721/JCSTCS.17309
004.652:004.912
natural language processing
named entity recognition
relation extraction
text analysis
classification
relational databases
model building
https://infocom.spbstu.ru/article/2024.82.9/
RAR
RUS
103-113
0000-0003-0885-1355
LLC “Predict space”
Aksenova
Lyubov
Novorossiysk, Russian Federation
0000-0001-5391-5229
LLC “Predict space”
Aksenov
Kirill
Novorossiysk, Russian Federation
LLC “Predict space”
Prysyazhnyuk
Anton
Novorossiysk, Russian Federation
Maykop State Technological University
Myasnikova
Viktoria
Maykop, Russian Federation
Bonch-Bruevich St. Petersburg State University of Telecommunications
Krasov
Andrei
St. Petersburg, Russian Federation
Development of an OCT data classification model for determining the presence and type of ophthalmic diseases
Optical Coherence Tomography (OCT) is an important tool in the diagnosis of common ophthalmological diseases, such as age-related macular degeneration and diabetic retinopathy. However, the processes of analyzing and interpreting OCT data are highly complex due to the need to process a large amount of data and the time spent on research, as well as the ophthalmologist's failure to recognize minor or early signs of the disease or rare pathologies. This paper proposes a comprehensive approach to the development of an OCT image analysis system based on deep neural networks. In particular, the performance of models based on four neural network architectures – ResNet50, VGG16, InceptionV4, and ResNet101 – was evaluated. The results show that the model based on the ResNet50 architecture achieves the highest proportion of correctly classified images. Furthermore, the integration of the developed model into a chatbot significantly reduces the time needed to interpret OCT images, which can contribute to increased availability of preliminary diagnostics and improved quality of medical services.
10.18721/JCSTCS.17310
303.732
artificial intelligence
ophthalmology
machine learning models
neural network architectures
convolutional neural networks
optical coherence tomography
chatbot
https://infocom.spbstu.ru/article/2024.82.10/
RAR
RUS
114-123
Special Technological Center Ltd.
Klimenko
Denis
d.klimenk0@yandex.ru
St. Petersburg, Russian Federation
Peter the Great St.Petersburg Polytechnic University
Nikitin
Aleksandr
nikitin@mail.spbstu.ru
Polytechnicheskaya, 29, St.Petersburg, 195251, Russia
Peter the Great St. Petersburg Polytechnic University
Stroganov
Alexander
lemyr103@gmail.com
St. Petersburg, Russian Federation
Monolithic integrated circuit of a four-channel switched filter bank for the centimeter band based on GaAs pНЕМТ technology
This article presents the results of the design of a four-channel switched filter bank for the centimeter band manufactured as a monolithic microwave integrated circuit based on domestic GaAs pHEMT technology. The switched filter bank includes a set of bandpass filters operating in four sub-bands of most of the C-, X- and Ku-bands, as well as broadband SP4T switches. The bandpass filters of the lower sub-bands are designed using lumped elements, and higher sub-bands filters are designed using microstrip hairpin resonators. The SP4T switch is based on SPST switches, each of which contains one series- and three parallel-connected field-effect transistors. The switched filter bank has following parameters: insertion losses of no more than 10.8 dB, stopband suppression of at least 43 dB at 30% offset or more from the passband center frequency, and a voltage standing wave ratio of no more than 1.8 in the passband.
10.18721/JCSTCS.17311
621.372.543
MMIC
switched filter bank
bandpass filter
SP4T
GaAs pHEMT
https://infocom.spbstu.ru/article/2024.82.11/
RAR
RUS
124-130
Peter the Great St. Petersburg Polytechnic University
Tran
Thanh Dat
thanhdat140495@gmail.com
St. Petersburg, Russian Federation
Enhanced frequency band wideband receiver using N-path Miller bandpass filter
This paper presents the design and modelling result of the wideband receiver topology with an enhanced frequency band using Miller N-path bandpass filter and harmonic-rejection mixing technique. The wideband receiver has the form of monolithic microwave integrated circuits (MMIC) based on domestic GaAs pHEMT technology. The receiver consists of low noise amplifier (LNA), commutated network and harmonic recombination circuit. When using inductors to compensate the parasitic capacitances, the bandwidth of the LNA significantly increases from 0–2.3 GHz to 0–5 GHz, thereby increasing the receiver frequency band to 0.3–3 GHz. The receiver achieves a gain of 15 dB, a noise figure of < 4 dB, an out-of-band IIP3 of +8 dBm, and a harmonic-rejection ratio at the third- and fifth-order local oscillator harmonics of > 50 dB.
10.18721/JCSTCS.17312
621.3.049.774.2
wideband receiver
multi-band receiver
wideband LNA
harmonic-rejection mixer
Miller N-path bandpass filter
MMIC
GaAs pHEMT
https://infocom.spbstu.ru/article/2024.82.12/
RAR
RUS
131-139
Yenuchenko
Mikhail
post@mixeme.ru
Peter the Great St. Petersburg Polytechnic University
Kvashina
Natalya
kvashina.nv@gmail.com
St. Petersburg, Russian Federation
Piatak
Ivan
i.m.piatak@gmail.com
Influence analysis of comparator parameters spread on decision accuracy in DAC self-calibration circuit
The article presents an analysis of the influence of the comparator parameter spread in the digital-to-analog converter (DAC) self-calibration circuit on the reduction of the conversion nonlinearity. DAC on sources with a switching-based self-calibration circuit is considered. The comparator response threshold value due to the spread of component parameters for 0.18 µm CMOS technology (HCMOS8D by “Mikron”) is estimated. The comparator response threshold values are obtained for three sizes of comparator components. Functional modeling of the switching calibration taking into account the finite threshold of element comparison showed that the choice of the sorting algorithm affects the reduction of the conversion nonlinearity. It should be noted that for the smallest comparator option, only quick sort can provide an improvement in the integral nonlinearity for all considered conditions. The optimal size of the comparator components is determined in terms of the efficiency of nonlinearity reduction. The quick sort algorithm shows the best results both in nonlinearity reduction and in the influence of the comparator switching threshold sign.
10.18721/JCSTCS.17313
621.3.049.774.2
digital-to-analog converter
current source
switching-based calibration
mismatch
comparator threshold
elements sorting
https://infocom.spbstu.ru/article/2024.82.13/
RAR
RUS
140-152
Peter the Great St. Petersburg Polytechnic University
Udalov
Pavel
pp_udalov@mail.ru
St. Petersburg, Russian Federation
Peter the Great St. Petersburg Polytechnic University
Lukin
Aleksei
lukin@compmechlab.com
St. Petersburg, Russian Federation
Peter the Great St. Petersburg Polytechnic University
Popov
Ivan
popov_ia@spbstu.ru
St. Petersburg, Russian Federation
0000-0003-3103-7060
Peter the Great St.Petersburg Polytechnic University
Loboda
Vera
Polytechnicheskaya, 29, St.Petersburg, 195251, Russian Federation
JSC “Institute of Nuclear Materials”
Varivcev
Artem
Zarechny, Russian Federation
JSC “Institute of Nuclear Materials”
Butakov
Denis
Zarechny, Russian Federation
Application of the perturbation method to construct a refined compact model of a thermoelectric element with temperature-dependent parameters
The paper presents an asymptotically substantiated compact model of the Peltier element. The problem of stationary temperature distribution in a one-dimensional thermoelectric medium with temperature-dependent physical parameters is considered. A direct asymptotic approximation is constructed under the assumption that the ratio of the temperature difference at the boundaries of the Peltier element to the mean absolute temperature of the module is a small value. Expressions for heat fluxes on the absorbing and radiating sides with a nonlinear dependence on the applied current and boundary temperatures are obtained. A method of synthesis based on the obtained solution of a compact system model of a thermoelectric module is proposed. A numerical example is used to compare the obtained model with the classical model with averaged material parameters. It is shown that the heat fluxes of the two models take different values at sufficiently large electric currents. Promising areas of using the proposed new analytical model of the Peltier element in industrial problems are discussed.
10.18721/JCSTCS.17314
537.322
Peltier battery
Matlab
Simscape
direct asymptotic expansion method
system-level modeling
reduced order modelling
https://infocom.spbstu.ru/article/2024.82.14/