Enabling Serverless Deployment of Large-Scale AI Workloads

Deployed AI platforms typically ship with bulky system architectures which present bottlenecks and a high risk of failure. A serverless deployment can mitigate these factors and provide a cost-effective, automatically scalable (up or down) and elastic real-time on-demand AI solution. However, deploying high complexity production workloads into serverless environments is far from trivial, e.g., due to factors such as minimal allowance for physical codebase size, low amount of runtime memory, lack of GPU support and a maximum runtime before termination via timeout. In this paper we propose a set of optimization techniques and show how these transform a codebase which was previously incompatible with a serverless deployment into one that can be successfully deployed in a serverless environment; without compromising capability or performance. The techniques are illustrated via worked examples that have been deployed live on rail data and realtime predictions on train movements on the UK rail network. The similarities of a serverless environment to other resource constrained environments (IoT, Mobile) means the techniques can be applied to a range of use cases.

Accurately predicting response times of service queries is necessary for deployments optimisation in the multi-tenant applications system. This task is particularly challenging owing to the fact that the mixes of tenants with different business scale and operating characteristics and the interaction among the concurrently running queries have a great impact on the response time of queries in the multi-tenant applications systems, and an accurate model needs to capture them. In this paper, our goal is to build such a performance model for the interactions of multi-tenant using an experiment-driven modelling approach. We use a Bayesian approach and build novel Gaussian models that take into account a variety of factors that influence the response time of each interaction that is sent from the different tenants in the multi-tenant environments. We experimentally demonstrate that our models are accurate and effective which have an average prediction error of 12.6% in the worst case.

When using an out-of-core indexing method to answer a query, it is generally assumed that the I/O cost dominates the overall query response time. Because of this, most research on indexing methods concentrate on reducing the sizes of indices. For bitmap indices, compression has been used for this purpose. However, in most cases, operations on these compressed bitmaps, mostly bitwise logical operations such as AND, OR, and NOT, spend more time in CPU than in I/O. To speedup these operations, a number of specialized bitmap compression schemes have been developed; the best known of which is the byte-aligned bitmap code (BBC). They are usually faster in performing logical operations than the general purpose compression schemes, but, the time spent in CPU still dominates the total query response time. To reduce the query response time, we designed a CPU-friendly scheme named the word-aligned hybrid (WAH) code. In this paper, we prove that the sizes of WAH compressed bitmap indices are about two words per row for large range of attributes. This size is smaller than typical sizes of commonly used indices, such as a B-tree. Therefore, WAH compressed indices are not only appropriate for low cardinality attributes but also for high cardinality attributes.In the worst case, the time to operate on compressed bitmaps is proportional to the total size of the bitmaps involved. The total size of the bitmaps required to answer a query on one attribute is proportional to the number of hits. These indicate that WAH compressed bitmap indices are optimal. To verify their effectiveness, we generated bitmap indices for four different datasets and measured the response time of many range queries. Tests confirm that sizes of compressed bitmap indices are indeed smaller than B-tree indices, and query processing with WAH compressed indices is much faster than with BBC compressed indices, projection indices and B-tree indices. In addition, we also verified that the average query response time is proportional to the index size. This indicates that the compressed bitmap indices are efficient for very large datasets.

<div>Connected vehicles and intelligent transportation systems are currently evolving into highly interconnected digital environments. Due to the interconnectivity of different systems and complex communication flows, a joint risk analysis for combining safety and security from a system perspective does not yet exist. We introduce a novel method for joint risk assessment in the automotive sector as a combination of the Diamond Model, Failure Mode and Effects Analysis (FMEA), and Factor Analysis of Information Risk (FAIR). These methods have been sequentially composed, which results in a comprehensive risk management approach to information security in an intelligent transport system (ITS). The Diamond Model serves to identify and structurally describe threats and scenarios, the widely accepted FMEA provides threat analysis by identifying possible error combinations, and FAIR provides a quantitative estimation of probabilities for the frequency and magnitude of risk events. We present the methodology and its step-by-step application on a practice-oriented automotive use case. As a result of this risk management approach, we can finally provide quantitative values from FAIR instead of a qualitative categorization, enabling a more accurate assessment of risks and prioritization of their mitigation. Simultaneously, the FMEA ensures complete risk identification at a component level. The approach is transparent, reusable, and can be adjusted to new estimations or insights easily and at any time, thus addressing the complexity and diversity of services in the transportation domain.</div>

Intelligent Traffic Systems (ITS), as integrated systems including control technologies, communication technologies, vehicle sensing and vehicle electronic technologies, have provided valuable solutions to the increasingly serious traffic problems. In the process of construction and operation of ITS, big data, especially multimedia data is produced at a rapid speed, which has made traffic information more and more complicated, causing traffic management facing new challenges. Hence, in order to achieve efficient management of all types of transportation resources and make better use of ITS, it is necessary and significant to study the architecture and performance of ITS in depth. Through dividing the system into different functional modules and assigning these modules to components in Performance Evaluation Process Algebra (PEPA), we can adopt a new method to realize the modeling and evaluating the working process of real-time positioning and route planning in ITS. Meanwhile, the fluid flow approximation is employed to conduct a performance analysis through PEPA models, guaranteeing that the response time, the maximum utilization and the throughput of the system can be achieved and analyzed.

Forecasting future trends constitutes a key process for supporting urban and territorial policy making in general. In this work, we explore how the domains of smart cities, smart transport, and smart energy will evolve until 2030 from a scientific and technological perspective, as a means to inform future policies for urban development in Europe. We started our work with an extensive review of recent and relevant research, covering policy and market reports, scientific journal articles, and other scientific publications. Then, a two-round Delphi survey with 120 field experts was conducted to assess the plausibility of the literature review findings to materialize until 2030. According to our empirical findings, there will be several speedy and structural changes in the three domains: we were able to identify a set of 18 statements that are highly probable to become reality in the next decade, whereas 17 statements were classified as plausible but not highly probable, and three statements raised controversiality. This work provides significant added value in supporting territorial policymakers' and stakeholders' decision making under uncertainty, as well as in designing highly relevant research agendas, attuned to contemporary and emerging trends.

Vehicular Ad Hoc Network (VANET) is an application of the general concept of Mobile Ad hoc Network (MANET) in the transportation domain. Playing a key-role in Intelligent Transportation Systems (ITS), VANETs are vulnerable to threats that may jeopardize the in-vehicle security, thus subsequently causing accidents. Moreover, on looking at the connected car paradigm that enables interaction between both in-vehicle services as well as devices, and services with outside ones, VANETs will have to inherit security risks of conventional information technologies (IT) systems. This makes it necessary to resort to approaches providing supports and services that are capable to tackle the resulting wide range of security risks. Many works have been dedicated to VANETs security. Mostly the focus has been on data security, relying on cryptographic centric tools in order to ensure privacy, anonymity, and data security. Contrary to this, we propose an approach that aims to proactively address issues beyond data security that could hinder VANETs to become a reality if not properly faced. Although inspired by honeypots used in wired networks, it differs because of the need for tools and methods that should be sufficiently efficient and flexible to ensure context-awareness of security measures.KeywordsCyber securityHoneypotVANETIntelligent Transport System

Data warehouse store historical records on which analysis queries are executed. Data warehouse are populated from the transaction database through a process called as Extract Transform and Load (ETL). Extract is to identify the transaction database records to be moved to the data warehouse. Transform is to make the records normalized to suit the data warehouse environment. Load is to actually store the records in the data warehouse. Today business wants to perform real time analysis of data. Availability of updated records in the data warehouse is necessary for real time analysis. This could be achieved through the process of near real time (ETL), i.e., to improve existing ETL process. Improvement to existing ETL could be achieved in many ways such as to increase the frequency of loads, to identify relevant changes that are required for analysis and move only those changes. Only through increasing the frequency of load would not be useful in many cases. Hence we should identify the changes and then move if they are useful for analysis making it smart ETL. In this paper we study three such techniques. First one is to create a replica of dimension table in data warehouse and move the changes to the replica this reduces the query response time. Second is to load the data in parallel so that loading time could be reduced. Third is to identify changes and trigger the loading process. We have used the first approach to create replica and loaded the data in parallel and observed that by loading in parallel not only does the loading time reduces but also the query response time. Then in the first approach is the query response time reduces to a certain limit methods suggest to move the data from replica dimensions to the original dimensions. Here we bring in right time trigger to move instead of just the query response time. We found that when we have a combined approach to query response time and right time trigger the number of moves are reduced.

Chapter 3 - An Agent Methodology for Processes, the Environment, and Services

The decreasing cost of computing makes it economically viable to reduce the response time of decision support queries by using parallel execution to exploit inexpensive resources. This goal poses the following query optimization problem: Minimize response time subject to constraints on throughput , which we motivate as the dual of the traditional DBMS problem. We address this novel problem in the context of Select-Project-Join queries by extending the execution space, cost model and search algorithm that are widely used in commercial DBMSs. We incorporate the sources and deterrents of parallelism in the traditional execution space. We show that a cost model can predict response time while accounting for the new aspects due to parallelism. We observe that the response time optimization metric violates a fundamental assumption in the dynamic programming algorithm that is the linchpin in the optimizers of most commercial DBMSs. We extend dynamic programming and show how optimization metrics which correctly predict response time may be designed.

A data warehouse is a central repository of historical data designed primarily to support analytical processing. These analytical queries are exploratory, long and complex in nature. Further, the rapid and continuous growth in the size of data warehouse increases the response times of such queries. Query response times need to be reduced in order to speedup decision making. This problem, being an NP-Complete problem, can be appropriately dealt with by using swarm intelligence techniques. One such technique, i.e. the set-based particle swarm optimization (SPSO), has been proposed to address this problem. Accordingly, a SPSO based view selection algorithm (SPSOVSA), which selects the Top-K views from a multidimensional lattice, is proposed. Experimental based comparison of SPSOVSA with the most fundamental view selection algorithm shows that SPSOVSA is able to select comparatively better quality Top-K views for materialization. The materialization of these selected views would improve the performance of analytical queries and lead to efficient decision making.

This paper looks at the processing of skyline queries on peer-to-peer (P2P) networks. We propose Skyframe, a framework for efficient skyline query processing in P2P systems, which addresses the challenges of quick response time, low network communication cost and query load balancing among peers. Skyframe consists of two querying methods: one is optimized for network communication while the other focuses on query response time. These methods are different in the way in which the query search space is defined. In particular, the first method uses a high dominating point that has a large dominating region to prune the search space to achieve a low cost in network communication. On the other hand, the second method relaxes the search space in order to allow parallel query processing to speed up query response. Skyframe achieves query load balancing by both query load conscious data space splitting/merging during the join/departure of nodes and dynamic load migration. We further show how to apply Skyframe to both the P2P systems supporting multi-dimensional indexing and the P2P systems supporting single-dimensional indexing. Finally, we have conducted extensive experiments on both real and synthetic data sets over two existing P2P systems: CAN (Ratnasamy in A scalable content-addressable network. In: Proceedings of SIGCOMM Conference, pp. 161---172, 2001) and BATON (Jagadish et al. in A balanced tree structure for peer-to-peer networks. In: Proceedings of VLDB Conference, pp. 661---672, 2005) to evaluate the effectiveness and scalability of Skyframe.

In the era of big data, computing exact answers to analytical queries becomes prohibitively expensive. This greatly increases the value of approaches that can compute efficiently approximate, but highly-accurate, answers to analytical queries. Alas, the state of the art still suffers from many shortcomings: Errors are still high unless large memory investments are made. Many important analytics tasks are not supported. Query response times are too long and thus approaches rely on parallel execution of queries atop large big data analytics clusters, in-situ or in the cloud, whose acquisition/use costs dearly. Hence, the following questions are crucial: Can we develop AQP engines that reduce response times by orders of magnitude, ensure high accuracy, and support most aggregate functions? With smaller memory footprints and small overheads to build the state upon which they are based? With this paper, we show that the answers to all questions above can be positive. The paper presents DBEst, a system based on Machine Learning models (regression models and probability density estimators). It will discuss its limitations, promises, and how it can complement existing systems. It will substantiate its advantages using queries and data from the TPC-DS benchmark and real-life datasets, compared against state of the art AQP engines.

Currently, a variety of information retrieval systems are available to potential users. These services are provided by commercial enterprises (such as Dow Jones [Dunn 1984] and The Source [Edelhart 1983]), while others are research efforts (the Boston Community Information System [Gifford1985]). While in many cases these systems are accessed from personal computers, typically no advantage is taken of the computing resources of those machines (such as local processing and storage). In this paper we explore the possibility of using the user's local storage capabilities to cache data at the user's site. This would improve the response time of user queries albeit at the cost of incurring the the overhead required in maintaining multiple copies. In order to reduce this overhead it may be appropriate to allow copies to diverge in a controlled fashion. This would not only make caching less costly, but would also make it possible to propagate updates to the copies more efficiently, e.g., when the system is lightly loaded, when communication tariffs are lower, or by batching together updates. Just as importantly, it also makes it possible to access the copies even when the communication lines or the central site are down. Thus, we introduce the notion of quasi-copies which embodies the ideas sketched above. We also define the types of deviations that seem useful, and discuss the available implementation strategies.

Modern cloud data warehouses are integral to processing heterogeneous query workloads, which range from quick online transactions to intensive ad-hoc queries and extract, transform, load (ETL) processes. The synchronization of heterogeneous workloads, particularly the blend of short and long-running queries, often degrades performance due to intricate concurrency controls and cooperative multi-tasking execution models. Additionally, the auto-scaling mechanisms for mixed workloads can lead to spikes in demand and underutilized resources, impacting both performance and cost-efficiency. This paper introduces the Flux, a cloud-native workload auto-scaling platform designed for Alibaba AnalyticDB, which implements a pioneering decoupled auto-scaling architecture. By separating the scaling mechanisms for short and long-running queries, Flux not only resolves performance bottlenecks but also harnesses the elasticity of serverless container instances for on-demand resource provisioning. Our extensive evaluations demonstrate Flux's superiority over traditional scaling methods, with up to a 75% reduction in query response time (RT), a 19.0% increase in resource utilization ratio, and a 77.8% decrease in cost overhead.