Cloud services are critical to society. However, their reliability is poorly understood. Towards solving the problem, we propose a standard repository for cloud uptime data. We populate this repository with the data we collect containing failure reports from users and operators of cloud services, web services, and online games. The multiple vantage points help reduce bias from individual users and operators. We compare our new data to existing failure data from the Failure Trace Archive and the Google cluster trace. We analyze the MTBF and MTTR, time patterns, failure severity, user-reported symptoms, and operator-reported symptoms of failures in the data we collect. We observe that high-level user facing services fail less often than low-level infrastructure services, likely due to them using fault-tolerance techniques. We use simulation-based experiments to demonstrate the impact of different failure traces on the performance of checkpointing and retry mechanisms. We release the data, and the analysis and simulation tools, as open-source artifacts available athttps://github.com/atlarge-research/cloud-uptime-archive.

People and businesses increasingly rely on public LLM services, such as ChatGPT, DALL·E, and Claude. Understanding their outages, and particularly measuring their failure-recovery processes, is becoming a stringent problem. However, only limited studies exist in this emerging area. Addressing this problem, in this work we conduct an empirical characterization of outages and failure-recovery in public LLM services. We collect and prepare datasets for 8 commonly used LLM services across 3 major LLM providers, including market-leads OpenAI and Anthropic. We conduct a detailed analysis of failure recovery statistical properties, temporal patterns, co-occurrence, and the impact range of outage-causing incidents. We make over 10 observations, among which: (1) Failures in OpenAI's ChatGPT take longer to resolve but occur less frequently than those in Anthropic's Claude;(2) OpenAI and Anthropic service failures exhibit strong weekly and monthly periodicity; and (3) OpenAI services offer better failure-isolation than Anthropic services. Our research explains LLM failure characteristics and thus enables optimization in building and using LLM systems. FAIR data and code are publicly available on https://zenodo.org/records/14018219 and https://github.com/atlarge-research/llm-service-analysis.

HPC datacenters offer a backbone to the modern digital society. Increasingly, they run Machine Learning (ML) jobs next to generic, compute-intensive workloads, supporting science, business, and other decision-making processes. However, understanding how ML jobs impact the operation of HPC datacenters, relative to generic jobs, remains desirable but understudied. In this work, we leverage long-term operational data, collected from a national-scale production HPC datacenter, and statistically compare how ML and generic jobs can impact the performance, failures, resource utilization, and energy consumption of HPC datacenters. Our study provides key insights, e.g., ML-related power usage causes GPU nodes to run into temperature limitations, median/mean runtime and failure rates are higher for ML jobs than for generic jobs, both ML and generic jobs exhibit highly variable arrival processes and resource demands, significant amounts of energy are spent on unsuccessfully terminating jobs, and concurrent jobs tend to terminate in the same state. We open-source our cleaned-up data traces on Zenodo (https://doi. org/10.5281/zenodo.13685426), and provide our analysis toolkit as software hosted on GitHub (https://github.com/atlarge-research/2024-icpads-hpc-workload-characterization). This study offers multiple benefits for data center administrators, who can improve operational efficiency, and for researchers, who can further improve system designs, scheduling techniques, etc.

The global e-commerce store needs to ensure compliance with various regulations at local, national, and international levels. One business use case is to identify face masks to avoid price gouging during times of high demand. In order to keep billions of items safe and legally compliant, it is important to ensure accurate classifications. Classification revisers aim to enhance classification accuracy by detecting and revising incorrect classifications. In this paper, we introduce this problem, along with appropriate online evaluation metrics for large-scale application scenarios. Our proposed method first learns neural network embedding from item textual features to define similar neighbors, and then simply uses these known classification results from k nearest neighbors to estimate an item’s class assignment. Our experiments demonstrate that it outperforms the state-of-the-art baseline approaches and its robustness to large scales. The proposed method can uniquely revise a significant number of classifications correctly, complementary to a multi-label classification system. The study indicates that our simple yet effective approach empowered by GPU computation is a viable solution of a commercial multi-label-classification revision problem.

In e-commerce, product classification is widely used for various purposes. Misclassifying products can cause compliance issues and hurt the company's reputation. To address this problem, we propose an automated system to proactively detect product misclassifications by overcoming several challenges. A large e-commerce retailer can sell billions of distinct products, on which many thousands of classification tasks are performed. At this massive scale, we need to quickly detect misclassifications under a limited budget. In this talk, we point out these challenges and show how we design our system to handle them. When evaluated on a set of Amazon's product classification data, at an overhead of <10% of the classification cost, our system automatically identified and corrected many misclassifications, which would take a human many thousand years to manually find and 14.6 years to manually review and correct if our system were not used.

To verify the authenticity of easily manipulated multimedia content, forensic researchers have proposed many techniques to estimate the processing history of given multimedia content. When multiple operations may be applied on multimedia content, a complete processing history would involve the information of not only what manipulation operations have been applied, but also in what order they were applied. However, there are few works considering the problem of detecting the order of operations. Moreover, due to the interplay among operations, the order of operations may not always be detectable. This leads to a fundamental question of when we can or cannot detect the order of operations. In this paper, we propose an information theoretical framework to answer this question. Specifically, we formulate the problem of detecting the order of operations into a multiple hypotheses testing problem. Then, we propose an information theoretical framework to characterize the relationship between the true hypothesis and the detected hypothesis. Under this framework, we propose a mutual information based criterion to determine the detectability of the order of operations. Furthermore, conditional fingerprints are defined in this framework to understand why the order of operations is not always detectable. The detection of the order of resizing and blurring is examined in this paper, where the order detection scheme has been proposed and the effectiveness of our framework has been demonstrated by simulations.

As it is more and more convenient to manipulate multimedia content, the authenticity of multimedia content becomes questionable. While there are many forensic techniques developed to identify the use of a single manipulation operation, a few has considered the cases where multiple operations may be involved. In these cases, investigators not only need to identify the use of each operation, but also need to detect the order of these operations. However, due to the interplay among operations, the order of operations may not always be detectable. This leads to a fundamental question of when we can and cannot detect the order of operations. In this paper, we formulate the problem of detecting the order of operations as a multiple hypotheses testing problem. Then, we propose an information theoretical framework to model the relationship between the detected hypothesis and the true hypothesis. Under this framework, we propose a mutual information-based criterion to obtain the best detector and use it to determine whether we can or cannot detect the order of operations based on certain set of features. A case study of detecting the order of resizing and blurring has been examined to demonstrate the effectiveness of the proposed framework and criteria. In addition, two known forensic problems are considered in the simulations to show that the results obtained from the proposed framework and criteria match those of the existing works.

While more and more forensic techniques have been proposed to detect the processing history of multimedia content, one starts to wonder if there exists a fundamental limit on the capability of forensics. In other words, besides keeping on searching what investigators can do, it is also important to find out the limit of their capability and what they cannot do. In this paper, we explore the fundamental limit of operation forensics by proposing an information theoretical framework. In particular, we consider a general forensic system of estimating operations' hypotheses based on extracted features from the multimedia content. In this system, forensicability is defined as the maximum forensic information that features contain about operations. Then, due to its conceptual similarity with mutual information in an information theory, forensicability is measured as the mutual information between features and operations' hypotheses. Such a measurement gives the error probability lower bound of all practical estimators, which use these features to detect the operations' hypotheses. Furthermore, it can determine the maximum number of hypotheses that we can theoretically detect. To demonstrate the effectiveness of our proposed information theoretical framework, we apply this framework on a forensic example of detecting the number of JPEG compressions based on normalized discrete cosine transform (DCT) coefficient histograms. We conclude that, when subband (2, 3) is used in detection and the size of the testing database is <;20000, the maximum number of JPEG compressions that we can expectedly perfectly detect using normalized DCT coefficient histogram features is four. Furthermore, we obtain the optimal strategies for investigators and forgers based on the fundamental measurement of forensicability.

Identifying a signal's origin and how it was acquired is an important forensic problem. While forensic techniques currently exist to determine a signal's acquisition history, these techniques do not account for the possibility that a signal could be compressively sensed. This is an important problem since compressive sensing techniques have seen increased popularity in recent years. In this paper, we propose a set of forensic techniques to identify signals acquired by compressive sensing. We do this by first identifying the fingerprints left in a signal by compressive sensing. We then propose two compressive sensing detection techniques that can operate on a broad class of signals. Since compressive sensing fingerprints can be confused with fingerprints left by traditional image compression techniques, we propose a forensic technique specifically designed to identify compressive sensing in digital images. In addition, we propose a technique to forensically estimate the number of compressive measurements used to acquire a signal. Through a series of experiments, we demonstrate that each of our proposed techniques can perform reliably under realistic conditions. Simulation results show that both our zero ratio detector and distribution-based detector yield perfect detections for all reasonable conditions that compressive sensing is used in applications, and the specific two-step detector for images can at least achieve probability of detection of 90% for probability of false alarm <;10%. In addition, our estimator for the number of compressive measurements can well reflect the real number.

A signal's compression history is of particular forensic significance because it contains important information about the origin and authenticity of a signal. Because of this, antiforensic techniques have been developed that allow a forger to conceal manipulation fingerprints. However, when antiforensic techniques are applied to multimedia content, distortion maybe introduced, or the data size may be increased. Furthermore,when compressing an antiforensically modified forgery, a tradeoff between the rate and distortion is introduced into the system. As a result, a forger must balance three factors, such as how much the fingerprints can be forensically concealed, the data rate, and the distortion, are interrelated to form a 3D tradeoff. In this paper, we characterize this tradeoff by defining concealability and using it to measure the effectiveness of an antiforensic attack. Then, to demonstrate this tradeoff in a realistic scenario, we examine the concealability-rate-distortion tradeoff in double JPEG compression antiforensics. To evaluate this tradeoff, we propose flexible antiforensic dither as an attack in which the forger can vary the strength of antiforensics. To reduce the time and computational complexity associated with decoding a JPEG file, applying antiforensics, and recompressing, we propose anantiforensic transcoder to efficiently complete these tasks in one step. Through simulation, two surprising results are revealed. One is that if a forger uses a lower quality factor in the second compression, applying antiforensics can both increase concealability and decrease the data rate. The other is that for any pairing of concealability and distortion values, achieved using a higher secondary quality factor, can also be achieved using a lower secondary quality factor at a lower data rate. As a result, the forger has an incentive to always recompress using a lower secondary quality factor.

Multimedia forensics concerns on extracting forensic information from suspicious multimedia contents. This information was embedded into the content inadvertently whenever an operation happened. Investigators may estimate the possible operations by obtaining features from the multimedia content and applying detection algorithms based on the statistics. While most existing works focus on improving detection performance and finding what more we can do, understanding the fundamental limit on the forensic information that we can obtain from the extracted features is also important. It enables us to understand the limit of forensicability. In this paper, we explore the fundamental limit of forensicability by introducing an information theoretical framework for multimedia forensics. We use mutual information as the measure of forensic information conveyed by features to investigators. To show the analytical process, we take the case of multiple JPEG compression forensics as an example. We claim that, under typical circumstances, the maximum number of compressions that we can detect by examine DCT coefficients is up to 4, in an expected sense. In addition, we also find the patterns of compression quality factors that contain the most and least forensic information.

In wireless sensor networks, a few sensor nodes end up being vulnerable to potentially rapid depletion of the battery reserves due to either their central location or just the traffic patterns generated by the application. In this paper, we use a new approach that balances the energy consumption at each of the nodes, thus increasing the functional lifetime of the network. We propose a new distributed dynamic topology control algorithm called Energy Balanced Topology Control (EBTC) which considers the actual energy consumed for each transmission and reception to achieve the goal of an increased functional lifetime. We analyze the algorithm's computational and communication complexity and show that it is equivalent or lower in complexity to other dynamic topology control algorithms. Using an empirical model of energy consumption, we show that the EBTC algorithm increases the lifetime of a wireless sensor network by over 40% compared to the best of previously known algorithms.

In this paper, we give a global perspective of multicast capacity and delay analysis in Mobile Ad Hoc Networks (MANETs). Specifically, we consider four node mobility models: (1) two-dimensional i.i.d. mobility, (2) two-dimensional hybrid random walk, (3) one-dimensional i.i.d. mobility, and (4) one-dimensional hybrid random walk. Two mobility time-scales are investigated in this paper: (i) fast mobility where node mobility is at the same time-scale as data transmissions and (ii) slow mobility where node mobility is assumed to occur at a much slower time-scale than data transmissions. Given a delay constraint$D$, we first characterize the optimal multicast capacity for each of the eight types of mobility models, and then we develop a scheme that can achieve a capacity-delay tradeoff close to the upper bound up to a logarithmic factor. In addition, we also study heterogeneous networks with infrastructure support.

Due to the ease with which digital multimedia content can be modified, a number of techniques to forensically detect forgeries have been developed. Meanwhile, anti-forensic operations have been developed to defeat forensic techniques. When anti-forensics is applied, a forger must balance between the amount that editing fingerprints have been concealed and the distortion introduced to the content. Additionally, the forger may compress the forgery for storage or transmission, which introduces a tradeoff between data rate and distortion. In this paper, we define a measure of an anti-forensic technique's effectiveness which we call concealability and examine the tradeoff between concealability, rate, and distortion. We then characterize the concealability-rate-distortion (C-R-D) surface for double JPEG compression anti-forensics. To do this, we propose a new technique known as flexible anti-forensic dither to hide double JPEG fingerprints. From our experiments, we identify two surprising results related to the C-R-D surface.

Due to the ease with which digital images can be forged, a great deal of work has been done in the field of digital image forensics. A particularly important problem is to forensically determine an image's acquisition and storage history. Recent research has shown that a new acquisition technique known as compressive sensing can be used to capture a digital image. Images acquired by compressive sensing can not be easily forensically distinguished from the images captured using standard digital cameras, nor from those which have undergone Discrete Wavelet Transform (DWT) based compressions. In this paper, we propose a two step detection scheme to identify the images acquired by compressive sensing. The first step identifies unaltered images captured using standard digital cameras and the second step separates compressively sensed images from ones compressed using DWT-based techniques. Our simulations show the first step of detection can achieve a probability of detection (Pd) of 100% for probability of false alarm (Pf) less than 4%, and the second step gets Pdnearly 90% with Pfof 10%.

Identifying a signal's origin and how it was acquired is an important problem for digital forensics. Recently, compressive sensing has achieved substantial attention due to its ability to accurately acquire sparse signals at rates below the Nyquist rate. The increased popularity of this signal acquisition technique gives rise to a new forensic problem: is it possible to distinguish signals that have been compressively sensed from traditionally sampled ones? In our previous work, we addressed this problem of differentiating between traditionally acquired and compressively sensed images. In this paper, we examine the problem of distinguishing traditionally sampled signals from compressively sensed ones for a broader class of signals. We categorize those compressive sensing applicable signals into two cases: sparse signals with noise and nearly sparse signals. For each category, we discuss the traces left in a signal by compressive sensing and propose a corresponding detection scheme. Experimental results show that both of our proposed detection schemes can be effectively used to distinguish compressively sensed signals from traditionally sensed signals.

Topology control algorithms allow each node in a wireless multi-hop network to adjust the power at which it makes its transmissions and choose the set of neighbors with which it communicates directly, while preserving global goals such as connectivity or coverage. This allows each node to conserve energy and contribute to increasing the lifetime of the network. Previous work on topology control has largely used an approach based on considering only the energy costs across links without considering the amount of energy available on a node. Further, previous work has largely used a static approach where the topology is determined at the beginning of the network's life and does not incorporate the varying rates of energy consumption at different nodes. In this paper, we address these weaknesses and introduce a new topology control algorithm that dynamically adapts to current energy levels at nodes. The algorithm, called Cooperative Topology Control with Adaptation (CTCA), employs a game-theoretic approach that maps the problem of maximizing the network's lifetime into an ordinal potential game. This allows a node running the CTCA algorithm to make a sacrifice by increasing its transmission power if it can help reduce energy consumption at another node with a smaller lifetime. We prove the existence of a Nash equilibrium for the game. Our simulation results indicate that the CTCA algorithm extends the life of a network by more than 50% compared to the best previouslyknown algorithm.

Each node in a wireless multi-hop network can adjust the power at which it makes its transmissions and thus change the topology of the network to save energy by choosing a smaller number of neighbors with which it communicates directly. The Directed Relative Neighborhood Graph (DRNG) algorithm is among the most popular, efficient and versatile topology control algorithms that reduces energy costs while preserving connectivity. In this paper, however, we identify two key limitations of DRNG and address them in a new power-aware distributed topology control algorithm called Inclusive DRNG. We show that the Inclusive DRNG algorithm can cut down the energy cost of executing the topology control algorithm itself to as low as one-third that of executing DRNG. We further show that Inclusive DRNG also generates a topology that achieves a significant improvement of over 25% in the average energy cost of communication across a path between any two nodes in realistic irregular radio environments.

In this paper, we give a global perspective of multicast capacity and delay analysis in Mobile Ad-hoc Networks (MANETs). Specifically, we consider two node mobility models: (1) two-dimensional i.i.d. mobility, (2) one-dimensional i.i.d. mobility. Two mobility time-scales are included in this paper: (i) Fast mobility where node mobility is at the same time-scale as data transmissions; (ii) Slow mobility where node mobility is assumed to occur at a much slower time-scale than data transmissions. Given a delay constraint D, we first characterize the optimal multicast capacity for each of the four mobility models, and then we develop a scheme that can achieve a capacity-delay tradeoff close to the upper bound up to a logarithmic factor. Our study can be further extended to two-dimensional/one-dimensional hybrid random walk fast/slow mobility models and heterogeneous networks.

BACKGROUND: An important aspect of protein design is the ability to predict changes in protein thermostability arising from single- or multi-site mutations. Protein thermostability is reflected in the change in free energy (DeltaDeltaG) of thermal denaturation. RESULTS: We have developed predictive software, Prethermut, based on machine learning methods, to predict the effect of single- or multi-site mutations on protein thermostability. The input vector of Prethermut is based on known structural changes and empirical measurements of changes in potential energy due to protein mutations. Using a 10-fold cross validation test on the M-dataset, consisting of 3366 mutants proteins from ProTherm, the classification accuracy of random forests and the regression accuracy of random forest regression were slightly better than support vector machines and support vector regression, whereas the overall accuracy of classification and the Pearson correlation coefficient of regression were 79.2% and 0.72, respectively. Prethermut performs better on proteins containing multi-site mutations than those with single mutations. CONCLUSIONS: The performance of Prethermut indicates that it is a useful tool for predicting changes in protein thermostability brought about by single- or multi-site mutations and will be valuable in the rational design of proteins.

In "area-sensing" applications of sensor networks, such as surveillance or target tracking, each sensor node has a sensing radius within which it can monitor events. Coverage problems in sensor networks have largely focused on such applications, where the goal of good coverage is one of ensuring that each point in the region of interest is within the sensing radius of at least one node. On the other hand, in "spot-sensing" applications, each node makes a measurement (such as temperature or humidity) at the precise location of the node and there is no concept of a sensing radius. In this paper, we introduce a new coverage problem that is more meaningful to spot-sensing applications. In such cases, good coverage usually implies even coverage across points in the region. We borrow from the field of economics and adapt a well-accepted measure of inequality, the Gini index, to develop a metric for the evenness of coverage by a sensor network. Based on mathematical results on the expected distances between neighboring nodes, we present a new distributed algorithm, called EvenCover, for each node to determine if and when it should sleep or sense. We prove that the expected Gini index is 1 - 1radic2 ap 0.293 when the spatial distribution of sensing nodes is given by a Poisson random process. On the other hand, when the sensing nodes are perfectly evenly distributed, we show that the Gini index has a lower bound of 0.2. These two results serve as points of reference to evaluate the coverage achieved by the EvenCover algorithm. We present a thorough simulation-based comparison of EvenCover against other distributed algorithms showing that it achieves better evenness and significantly increased lifetime. In addition, we discover that evenness of coverage permits a graceful degradation of the network as nodes exhaust their energy resources.