In an inference service system, model selection and scheduling (MS&S) schemes map inference queries to trained machine learning (ML) models, hosted on a finite set of workers, to solicit accurate predictions within strict latency targets. MS&S is challenged by both varying query load and stochastic query inter-arrival patterns; however, state-of-the-art MS&S approaches conservatively account for load exclusively.

New ML models are often derived from existing ones (e.g., through fine-tuning, quantization or distillation), forming an ecosystem where models are related to each other and can share structure or even parameter values. Managing such a large and evolving ecosystem of model derivatives is challenging. For instance, the overhead of storing all such models is high, and models may inherit bugs from related models, complicating error attribution and debugging. In this paper, we propose a model versioning and management system called MGit that makes it easier to store, test, update, and collaborate on related models. MGit introduces a lineage graph that records the relationships between models, optimizations to efficiently store model parameters, and abstractions over this lineage graph that facilitate model testing, updating and collaboration. We find that MGit works well in practice: MGit is able to reduce model storage footprint by up to 7$\times$. Additionally, in a user study with 20 ML practitioners, users complete a model updating task 3$\times$ faster on average with MGit.

The Sandia Frequency Shift (SFS) method of anti-islanding protection is considered to be the most effective scheme built into grid-connected inverters. What makes it so effective is the positive feedback loop that it creates when the inverter’s connection to the bulk power system is lost. However, this same positive feedback can have negative impacts on the stability of the inverter even when it is connected to the bulk power system. We were able to show that SFS can impact the small signal stability of a grid-connected inverter by impacting the system’s eigenvalues and can reduce the inverter’s power transfer stability limit. The extent of these impacts depends on the SFS formulation used and whether the inverter is connected to a "weak grid". This impact may be even greater in an intentionally islanded microgrid.

Abstract A rigorous formalization of desired system requirements is indispensable when performing any verification task. This often limits the application of verification techniques, as writing formal specifications is an error-prone and time-consuming manual task. To facilitate this, we present , a framework for applying Large Language Models (LLMs) to derive formal specifications (in temporal logics) from unstructured natural language. In particular, we introduce a new methodology to detect and resolve the inherent ambiguity of system requirements in natural language: we utilize LLMs to map subformulas of the formalization back to the corresponding natural language fragments of the input. Users iteratively add, delete, and edit these sub-translations to amend erroneous formalizations, which is easier than manually redrafting the entire formalization. The framework is agnostic to specific application domains and can be extended to similar specification languages and new neural models. We perform a user study to obtain a challenging dataset, which we use to run experiments on the quality of translations. We provide an open-source implementation, including a web-based frontend.

Out-of-order parallel discrete event simulation (PDES) has been shown to be very effective in speeding up system design by utilizing parallel processors on multi- and many-core hosts. As the number of threads in the design model grows larger, however, the original scheduling approach does not scale. In this work, we analyze the out-of-order scheduler and identify a bottleneck with quadratic complexity in event prediction. We propose a more efficient lazy strategy based on defining trees and a schedule bypass with O(m log2m) complexity which shows sustained and improved performance gains in simulation of SystemC models with many processes. For models containing over 1000 processes, experimental results show simulation run time speedups of up to 90× using lazy event prediction against the original out-of-order PDES approach.

This paper proposes a general framework that allows to identify a video in real time using perceptual image hashing algorithms. In order to evaluate the versatility and performance of the framework, it was coupled for a use case about ads tv monitoring. Four Perceptual Image Hashing (PIH) algorithms were subject to a benchmarking process in order to identify the best one for the use case. This process was focused on analyze differences in terms of discriminability (D), robustness (R), time processing (Tp) and efficiency (E). A truth table was used to obtain information about discriminability and robustness, while processing time was directly measured. An efficiency metric based on time processing and identification capacity was proposed. In general terms, DHASH and PHASH algorithms have higher identification capacities than AHASH and WHASH in order to identify a video using only one frame. Moreover, a progressive decrease in robustness with the increment of the Hamming distance is observed in all cases. However, in a specific case of tv monitoring where speed is critical, the processing time becomes the most discriminatory parameter for the selection of the algorithm. So, for this case, a particular type of PIH (Average Hash) is highlighted as the most efficient one among other techniques, reaching an accuracy of 100% and frame rates on processing average of 108 fps with a Hamming Distance of 1. At the end, the proposed framework has remarkable identification skills, and presents an efficient search. Furthermore, presents the steps to select the best algorithm and its more adequate parameters, according to the requirements of each particular case.