Euiseong Seo’s research while affiliated with Sungkyunkwan University and other places

... However, implementing high-accuracy but resource-intensive ML models like convolutional neural networks (CNNs) poses challenges due to limited computational and memory resources [23]. Recent research addresses these challenges through two main aspects: optimizing ML models for resource efficiency and enhancing energy management [8,9,11,15,18]. ...

Reference:

E-QUARTIC: Energy Efficient Edge Ensemble of Convolutional Neural Networks for Resource-Optimized Learning

... We observed such drift and will address this challenge in future work. Power/Energy Characterization and Optimization: Researchers used the tools and methods above to characterize the energy efficiency of critical workloads and primitives in AI and HPC running on different scales [38]- [44], and to study the efficiency of the latest innovations in GPUs and other accelerators [45], [46]. Prior work also investigated the impact of frequency capping, power capping, DVFS, and input data composition on energy efficiency [24], [47]- [51]. ...

Reference:

Methodology for Fine-Grain GPU Power Visibility and Insights

... Leveraging heterogeneous GPUs is challenging due to their varying hardware properties (e.g., memory size, computation capability). To split data, prior work has looked at scaling the batch size in data parallelism for each GPU [17,24]. To split the model, prior work has explored unevenly distributing model layers in pipeline parallelism across GPUs [15,42,45]. ...

Reference:

HeterMoE: Efficient Training of Mixture-of-Experts Models on Heterogeneous GPUs

... Traditionally, in High-Performance Computing (HPC), CNN inference is performed solely on GPUs. Therefore, multiple works only focus on power efficiency for inference on GPUs, such as [18] and [22]. PELSI, on the other hand, focuses on HMPSoCs where embedded CPUs and GPUs are comparable in performance, and both are used for inference to maximize efficiency. ...

Reference:

PELSI: Power-Efficient Layer-Switched Inference

... PMem's low bandwidth (compared to DRAM) has been researched extensively [4,11,20,23,29,44]. Plush speci cally addresses issues raised by Gugnani et al. [15] and Woo et al. [44]. ...

Reference:

Plush: A Write-Optimized Persistent Log-Structured Hash-Table

... 7. Disable Non-Uniform Memory Access (NUMA) Memory Balancing Automatic migration of memory pages between NUMA nodes would cause memory page faults and latency spikes similar to that of disk-based swapping, for application tasks that used the affected memory. We therefore disabled this type of memory balancing feature [57,58]. 8. Disable Mitigations for CPU Vulnerabilities Mitigations against CPU vulnerabilities such as Spectre, Retbleed and Downfall to name a few, should be turned off, as those mitigations can have an unknown negative effect on certain types of application tasks. ...

Reference:

Os noise mitigations for benchmarking web browser execution environment performance

... Environmental adaptation and customization are critical factors in developing robust fitness tracking systems that can generalize across diverse workout scenarios. The proposed system employs techniques such as transfer learning [12], domain adaptation [13], and user profiling [14] to enable models to adapt to varying lighting conditions, backgrounds, and user preferences. These techniques have been successfully applied in computer vision tasks [15], allowing models to leverage knowledge from related domains and user-specific data to improve performance in new scenarios. ...

Reference:

Adaptive Exercise Meticulousness in Pose Detection and Monitoring via Machine Learning

... Idempotent GPU kernels promise to produce the same output upon re-execution, without causing any side effects, no matter how many times it is interrupted at any point. Such idempotence property has been shown to be effective in overcoming the performance bottlenecks of many aspects of GPU systems, including fault tolerance [37,76], task preemption [22,36,54], and memory persistence [44,75]. For instance, prior work shows that idempotence-based optimizations can significantly reduce the checkpointing overheads of fault-tolerant systems to less than 1% [37]. ...

Reference:

Microsecond-scale Dynamic Validation of Idempotency for GPU Kernels

... To fulfill user expectations of seamless and rapid application relaunch, mobile systems preserve all execution-related data (called anonymous data in Linux [4]), such as stack and heap, in main memory. This practice, known as keeping applications alive in the background [1,[5][6][7][8], enables faster relaunches. However, it also results in significant main memory capacity requirements for each application. ...

Reference:

Ariadne: A Hotness-Aware and Size-Adaptive Compressed Swap Technique for Fast Application Relaunch and Reduced CPU Usage on Mobile Devices

... ETC incurs an inter-phase cost both before and after adjusting the TLP, as it requires a period to detect changes in locality loss and reach the optimal TLP during runtime. Consequently, the best thread throttling strategy chosen statically for each phase outperforms dynamic thread throttling, as the static scheme does not incur inter-phase cost [16,17]. Several other studies have also proposed page prefetch and eviction policies to reduce the overhead caused by UVM oversubscription [1-4, 8-10, 18-20]. ...

Reference:

GPU thread throttling for page-level thrashing reduction via static analysis