Donghyun Kang

As mobile devices evolve to handle increasingly complex and I/O intensive applications, the demand for high-performance storage devices increases. To meet increasing demands, the zoned namespace (ZNS) interface has emerged as a promising alternative and is enabled on traditional UFS storage devices on mobile devices, called ZUFS. In this paper, we first explore emerging TLC-based ZUFS storage devices built on high-density TLC NAND cells due to the physical form-factor constraints of mobile devices. Then, we propose PoZS that enhances overall performance by utilizing the asymmetric bit performance of TLC NAND cells through vertical programming ordering (VPO) and classification of the physical zones within the TLC-based ZUFS into three groups, namely LSB-Z, CSB-Z, and MSB-Z. For evaluation, we implement PoZS in the NVMeVirt emulator by enabling the zoned namespace interface, and compare it with the baseline. To validate the robustness and effectiveness of PoZS, we conducted extensive evaluations across a broad spectrum of storage scenarios, including various mobile and server workloads. Experimental results clearly confirm that PoZS consistently outperforms baselines, achieving significant improvements in I/O throughput while substantially increasing LSB pages; it achieves a performance improvement of 117% in a real-world mobile workload. These findings show that PoZS is highly effective under high-density storage constraints and meets the performance demands of next-generation mobile devices.

The memory demand of modern applications has been rapidly increasing with the continuous growth of data volume across industrial and academic domains. As a result, computing devices (i.e., IoT devices, smartphones, and tablets) often experience memory shortages that degrade system performance and quality of service by wasting CPU cycles and energy. Thus, most operating systems rely on the swap mechanism to mitigate the memory shortage situation in advance, even if the swap memory fragmentation problem occurs over time. In this paper, we analyze the fragmentation behavior of the swap memory space within storage devices over time and demonstrate that the latency of swap operations increases significantly under aged conditions. We also propose a new extension of the traditional swap mechanism, called VSwap, that mitigates the swap memory fragmentation problem in advance by introducing two core techniques, virtual migration and address remapping. In VSwap, virtual migration gathers valid swap pages scattered across multiple clusters into contiguous regions within the swap memory space, while address remapping updates the corresponding page table entries to preserve consistency after migration. For experiments, we enable VSwap on the traditional swap mechanism (i.e., kswapd) by implementing it with simple code modifications. To confirm the effectiveness of VSwap, we performed a comprehensive evaluation based on various workloads. Our evaluation results confirm that VSwap is more useful and highly valuable than the original swap mechanism. In particular, VSwap improves the overall performance up to 48.18% by harvesting available swap memory space in advance with negligible overhead; it performs close to the ideal performance.

This study addresses the challenge of meeting deadlines in real-time applications such as video streaming and online gaming, in which data block delivery is time sensitive. Traditional loss-recovery solutions such as retransmission are inadequate, particularly in multi-path networks with long round-trip times, resulting in delays and missed deadlines. The proposed deadline-aware redundant transmission (DART) approach proactively transmits redundant packets across multiple paths, considering the next transmission time and one-way delay on each path to meet deadlines with minimal overhead. It also addresses deadlocks caused by insufficient network capacity by prioritizing the transmission of the smallest data block in the queue. DART, which operates independently of existing schedulers, simplifies implementation and reduces disturbances to existing schedulers. Simulation results demonstrate that DART increases the block transmission success rates by 5% to 10%, depending on the network conditions, while requiring minimal overheads.

This paper aims to address sensor-related challenges in simultaneous localization and mapping (SLAM) systems, specifically within the open-source Google Cartographer project, which implements graph-based SLAM. The primary problem tackled is the adaptability and functionality of SLAM systems in diverse robotic applications. To solve this, we developed a novel SLAM framework that integrates five additional functionalities into the existing Google Cartographer and Robot Operating System (ROS). These innovations include an inertial data generation system and a sensor data preprocessing system to mitigate issues arising from various sensor configurations. Additionally, the framework enhances system utility through real-time 3D topographic mapping, multi-node SLAM capabilities, and elliptical sensor data filtering. The average execution times for sensor data preprocessing and virtual inertial data generation are 0.55 s and 0.15 milliseconds, indicating a low computational overhead. Elliptical filtering has nearly the same execution speed as the existing filtering scheme.

With the evolution of cellular networks and wireless-local-area-network-based communication technologies, services for smart device users have appeared. With the popularity of 4G and 5G, smart device users can now consume larger bandwidths than before. Consequently, the demand for various services, such as streaming, online games, and video conferences, has increased. For improved quality of experience (QoE), streaming services utilize adaptive bitrate (ABR) algorithms to handle network bandwidth variations. ABR algorithms use network bandwidth history for future network bandwidth prediction, allowing them to perform efficiently when network bandwidth fluctuations are minor. However, in environments with frequent network bandwidth changes, such as wireless networks, the QoE of video streaming often degrades because of inaccurate predictions of future network bandwidth. To address this issue, we utilize the gated recurrent unit, a time series prediction model, to predict the network bandwidth accurately. We then propose a buffer-based ABR streaming technique that selects optimized video-quality settings on the basis of the predicted bandwidth. The proposed algorithm was evaluated on a dataset provided by Zeondo by categorizing instances of user mobility into walking, bus, and train scenarios. The proposed algorithm improved the QoE by approximately 11% compared with the existing buffer-based ABR algorithm in various environments.

Today, energy and performance efficiency have become a crucial factor in modern computing environments, such as high-end mobile devices, desktops, and enterprise servers, because data volumes in cloud datacenters increase exponentially. Unfortunately, many researchers and engineers neglect the power consumption and internal performance incurred by storage devices. In this paper, we present a renewable-zoned namespace (ReZNS), an energy and performance-optimal mechanism based on emerging ZNS SSDs. Specifically, ReZNS recycles the remaining capacity of zones that are no longer used by adding a renewable concept into the mapping mechanism. We implemented a prototype of ReZNS based on NVMeVirt and performed comprehensive experiments with diverse workloads from synthetic to real-world workloads to quantitatively confirm power and performance benefits. Our evaluation results present that ReZNS improves overall performance by up to 60% and the total power consumption by up to 3% relative to the baseline on ZNS SSD. We believe ReZNS creates new opportunities to prolong the lifespan of various consumer electronics, such as TV, AV, and mobile devices, because storage devices play a crucial role in their replacement cycle.

In deep learning (DL) frameworks, a checkpoint operation is widely used to store intermediate variable values (e.g., weights, biases, and gradients) on storage media. This operation helps to reduce the recovery time of running a machine learning (ML) model after sudden power failures or random crashes. However, the checkpoint operation can stall the overall training step of the running model and waste expensive hardware resources by leaving the GPU in idle sleep during the checkpoint operation. In addition, the completion time of the checkpoint operation is unpredictable in cloud server environments (e.g., AWS and Azure) because excessive I/O operations issued by other running applications interfere with the checkpoint operations in the storage stacks. To efficiently address the above two problems, we carefully designed Check-QZP, which reduces the amount of data required for checkpoint operations and parallelizes executions on the CPU and GPU by understanding the internal behaviors of the training step. For the evaluation, we implemented Check-QZP and compared it with the traditional approach in real-world multi-tenant scenarios. In the evaluation, Check-QZP outperformed the baseline in all cases in terms of the overall checkpoint time and the amount of data generated by the checkpoint operations, reducing them by up to 87.5% and 99.8%, respectively. In addition, Check-QZP achieved superior training speeds compared to the baseline.

Recently, the technologies of on-device AI have been accelerated with the development of new hardware and software platforms. Therefore, many researchers and engineers focus on how to enable ML technologies on mobile devices with limited hardware resources. In this paper, we revisit on-device ML designed to support ML technologies on mobile devices and describe the three challenges when using on-device ML in detail. Then, we propose a new data management policy, called Overlay-ML, which efficiently solves two challenges that we discovered. Especially, we designed Overlay-ML to work in the application space with two key ideas. The first key idea is to extend the limited memory space using the usable space of the underlying storage device. The second key idea is to provide data transparency, which hides where the data is stored so that running ML models think the data is stored in the same place. For evaluation, we implemented an image detection application based on TensorFlow Lite which is a well-known on-device ML framework, and modified it to enable the features of Overlay-ML. All evaluation was performed on two state-of-the-art smartphones that are high-end embedded devices. Our evaluation results clearly show Overlay-ML can effectively prevent unexpected termination by Android OS and present a good loss value in real-world workload.

Recently, realistic services like virtual reality and augmented reality have gained popularity. These realistic services require deterministic transmission with end-to-end low latency and high reliability for practical applications. However, for these real-time services to be deterministic, the network core should provide the requisite level of network. To deliver differentiated services to each real-time service, network service providers can classify applications based on traffic. However, due to the presence of personal information in headers, application classification based on encrypted application data is necessary. Initially, we collected application traffic from four well-known applications and preprocessed this data to extract encrypted application data and convert it into model input. We proposed a lightweight transformer model consisting of an encoder, a global average pooling layer, and a dense layer to categorize applications based on the encrypted payload in a packet. To enhance the performance of the proposed model, we determined hyperparameters using several performance evaluations. We evaluated performance with 1D-CNN and ET-BERT. The proposed transformer model demonstrated good performance in the performance evaluation, with a classification accuracy and F1 score of 96% and 95%, respectively. The time complexity of the proposed transformer model was higher than that of 1D-CNN but performed better in application classification. The proposed transformer model had lower time complexity and higher classification performance than ET-BERT.