This is the official repository for paper "Performance Modeling of Data Storage Systems using Generative Models" in IEEE Access [journal] [arxiv].
High-precision systems modeling is one of the main areas of industrial data analysis. Models of systems, their digital twins, are used to predict their behavior under various conditions. In this study, we developed several models of a storage system using machine learning-based 8000 generative models to predict performance metrics such as IOPS and latency. The models achieve prediction errors ranging from 4%–10% for IOPS and 3%–16% for latency and demonstrate high correlation (up to 0.99) with observed data. By leveraging Little’s law for validation, these models provide reliable performance estimates. Our results outperform conventional regression methods, offering a vendor-agnostic approach for simulating data storage system behavior. These findings have significant applications for predictive maintenance, performance optimization, and uncertainty estimation in storage system design.
If you use the scripts or datasets in a scientific publication, we would appreciate citations to the paper:
A. R. Al-Maeeni, A. Temirkhanov, A. Ryzhikov and M. Hushchyn, "Performance Modeling of Data Storage Systems Using Generative Models," in IEEE Access, vol. 13, pp. 49643-49658, 2025, doi: 10.1109/ACCESS.2025.3552409
or using BibTeX:
@ARTICLE{10930879,
author={Al-Maeeni, Abdalaziz R. and Temirkhanov, Aziz and Ryzhikov, Artem and Hushchyn, Mikhail},
journal={IEEE Access},
title={Performance Modeling of Data Storage Systems Using Generative Models},
year={2025},
volume={13},
number={},
pages={49643-49658},
doi={10.1109/ACCESS.2025.3552409}}
All datasets used in the paper are in dataset.tar.gz archive with the following structure:
dataset
├── cache
│ ├── cache.csv # Full dataset for the cache under the random dataload.
│ ├── test_cache.csv # Test sample from the full cache dataset.
│ └── train_cache.csv # Train sample from the full cache dataset.
└── pools
├── hdd_sequential.csv # Full dataset for the HDD pool under the sequential dataload.
├── ssd_random.csv # Full dataset for the SSD pool under the random dataload.
├── ssd_sequential.csv # Full dataset for the SSD pool under the sequential dataload.
├── test_hdd_sequential.csv # Test sample from the full HDD pool dataset.
├── test_ssd_random.csv # Test sample from the full SSD pool dataset.
├── test_ssd_sequential.csv # Test sample from the full SSD pool dataset.
├── train_hdd_sequential.csv # Train sample from the full HDD pool dataset.
├── train_ssd_random.csv # Train sample from the full SSD pool dataset.
└── train_ssd_sequential.csv # Train sample from the full SSD pool dataset.iops: (target) the number of input and output (IO) operations per second;
lat: (target) the average latencies of input and output operations;
block_size: the size of one data block processed per one IO operation;
n_jobs: the number of processes in parallel used to process a IO dataload;
iodepth: the IO queue size per job;
read_fraction: the fraction of read operations in the dataload;
load_type: sequential or random dataload;
io_type: read or write operations;
raid: RAID scheme used in HDD or SSD pools. It is defined by the number of data (K) and parity (M) blocks in K+M format;
n_disks: the number of HDD or SSD disks used in the pools;
device_type: name of a device. Can be ignored;
offset: constant value. Can be ignored;
id: IDs of dataloads in the samples.
The dataset is also available on Kaggle.
# clone the project to your local machine
git clone https://github.com/HSE-LAMBDA/digital-twin.git
# install package
cd digital-twin && pip install -e .
.
├── digital_twin # Directory with a library source code
│ ├── models #
| | ├── density estimation #
| | | ├── gmm.py #
| | | ├── knn.py #
| | | ├── regressor.py #
| | | ├── grouper.py #
| | ├── norm_flow # Directory for normalized-flow-based digital twin model
| | | ├── utils #
| | | ├── model.py #
| ├── performance metrics #
| | ├── fd.py # Frechet Distance metric
| | ├── misc.py #
| | ├── mmd.py # Maximum Mean Discrepancy metric
| ├── visualization #
| | ├── plot.py #
| ├── data.py #
├── model_checkpoints #
├── results # Directory with experiments results
├── scripts # Scripts to reproduce experiments
| ├── calculate_summary_stats.py # This script calculates metrics and plot prediction plots based on prediction csv files
| ├── denesity_estimation_experiment.py #
| ├── generate_train_test_split.py #
| ├── knn_experiment.py #
| ├── nf_experiment.py # This script is used for experiment with normalized flow