Large Language Models (LLMs) achieve impressive reasoning capabilities at the cost of substantial inference overhead, posing substantial deployment challenges. Although distilled Small Language Models (SLMs) significantly enhance efficiency, their performance suffers as they fail to follow LLMs' reasoning paths. Luckily, we reveal that only a small fraction of tokens genuinely diverge reasoning paths between LLMs and SLMs. Most generated tokens are either identical or exhibit neutral differences, such as minor variations in abbreviations or expressions. Leveraging this insight, we introduce **Roads to Rome (R2R)**, a neural token router that selectively utilizes LLMs only for these critical, path-divergent tokens, while leaving the majority of token generation to the SLM. We also develop an automatic data generation pipeline that identifies divergent tokens and generates token-level routing labels to train the lightweight router. We apply R2R to combine R1-1.5B and R1-32B models from the DeepSeek family, and evaluate on challenging math, coding, and QA benchmarks. With an average activated parameter size of 5.6B, R2R surpasses the average accuracy of R1-7B by 1.6×, outperforming even the R1-14B model. Compared to R1-32B, it delivers a 2.8× wall-clock speedup with comparable performance, advancing the Pareto frontier of test-time scaling efficiency.

This paper proposes a point cloud background filtering method and explores applications that integrate LiDAR technology into roadside sensor networks, addressing challenges in handling large volumes of sparse and unstructured 3D data. Traditional methods, which classify background points based on descriptive statistics over many frames, are computationally inefficient for roadside LiDAR surveillance. This study introduces a novel approach using hash-based transformation combined with probabilistic Gaussian Mixture Model (GMM) techniques, enhancing the efficiency of high-dimensional multivariate modeling of LiDAR data. This Hash-based Gaussian Mixture Modeling (HGMM) optimizes feature selection and Gaussian components using AIC and BIC scores, mitigating computational burdens and parameter sensitivity. Unlike Cartesian coordinate-based techniques, HGMM processes LiDAR data in Spherical coordinates, preserving meaningful patterns and structures. For infrastructure LiDAR, object detection only pertains to a small amount of data in a fixed environment, allowing background reduction modeling to significantly enhance data chain efficiency by transitioning only a tiny portion of the foreground LiDAR point clouds. The method was tested on diverse LiDAR datasets, including the New Brunswick DataCity Testbed dataset, Transportation Forecasting Competition (TRANSFOR 24) Dataset, DAIR-V2X, and A9-Dataset, demonstrating its adaptability and efficiency across different scenarios. This approach enhances LiDAR sensors’ utility in supporting AI-enhanced decision-making processes and enriching the knowledge base of incorporating LiDAR sensors into current and prospective traffic management strategies.

Wider SIMD units suffer from low scalability of gather/scatter instructions that appear in sparse matrix calculations. We address this problem with an out-of-step pipeline which tolerates bank conflicts of a multibank L1D by allowing element operations of SIMD instructions to proceed out of step with each other. We evaluated it with a sparse matrix-vector product kernel for matrices from HPCG and SuiteSparse Matrix Collection. The results show that, for the SIMD width of 1024 bit, it achieves 1.91 times improvement over a model of a conventional pipeline.

Semantic segmentation (SS) is an important perception manner for self-driving cars and robotics, which classifies each pixel into a pre-determined class. The widely-used cross entropy (CE) loss-based deep networks has achieved significant progress w.r.t. the mean Intersection-over Union (mIoU). However, the cross entropy loss can not take the different importance of each class in an self-driving system into account. For example, pedestrians in the image should be much more important than the surrounding buildings when make a decisions in the driving, so their segmentation results are expected to be as accurate as possible. In this paper, we propose to incorporate the importance-aware inter-class correlation in a Wasserstein training framework by configuring its ground distance matrix. The ground distance matrix can be pre-defined following a priori in a specific task, and the previous importance-ignored methods can be the particular cases. From an optimization perspective, we also extend our ground metric to a linear, convex or concave increasing function w.r.t. pre-defined ground distance. We evaluate our method on CamVid and Cityscapes datasets with different backbones (SegNet, ENet, FCN and Deeplab) in a plug and play fashion. In our extenssive experiments, Wasserstein loss demonstrates superior segmentation performance on the predefined critical classes for safe-driving.

The extraction of urban impervious surface information plays a key role in the studies of urbanization and its related environmental issues. Optical and SAR remote sensing provides complementary information to improve the accuracy of impervious mapping. However, the fusing of information acquired by different sensors is challenging. Optical and SAR features have distinct characteristics, and require different classification strategy and classification types. In this study, a strategy of fusing multi-spectral optical and polarimetric SAR data at decision-level is proposed. Features are extracted from optical and SAR data, then staked auto-encoder is applied to achieve the land use and land cover classification separately. D-S evidence theory is used to fuse the classification result and the imperious surface map is derived. The experiment was conducted in a highly complex urban area of Hong Kong and the results proves the soundness of the method.

Emerging Internet of Things(IoT) applications are moving from silo and small scale sensor data sharing to composite and large scale ones. With the rapid growth of application scale, IoT applications is going to leverage cloud infrastructure for scalable solutions and real-time services. Thus large volumes of heterogeneous and high frequency sensor data are fed into IoT cloud services for real-time actionable insight, which raises great challenges of performance and adaptability on cloud solutions. In this paper, we propose a streaming based processing infrastructure for high throughput and low latency IoT real-time analytics services. A data adaptive mechanism is also introduced for heterogeneous data stream integration, interpreting and processing with application logics, as well as context stream. We implemented the proposed mechanisms with spark streaming, and deployed real time IoT analytics service in cloud. Experiment results show that the service has a good scalability and high throughput for IoT data analytics.

Push notification is an important approach to distribute interesting information to users timely. With the fast development of mobile devices and mobile applications, push notification is getting more and more popular. The convergence of mobile and IoT also bring new challenges on how the system can handle the mixed push channels designed for M2M communication and human interaction, and enable the effective interaction with both human and IoT devices involved. IoT devices may push notifications of sensor data in a high frequency. To enable push notification for both of mobile devices and IoT devices, the push notification system is required to achieve high throughput to handle the frequent notifications, and support content matching to filter out the undesired notifications. To enable effective push notification with both human and IoT devices involved, the system is required to understand the users' interests for notifications with the IoT devices providing the users' contexts. That is to say users need an intelligent push notification system. In this paper we propose a high performance context-aware push notification system for the converged mobile and IoT messaging. We designed high performance content matching engine as the core to enable efficient message dispatching for push notification according to highly personalized interest to ensure IoT messages to be involved in push notification. A user's interest of notifications is highly related with his context. Therefore, based on the content-matching engine, a framework for efficient context information fusion is built to support various types of context-aware push notification, towards intelligent push notification. Also we designed shared connection scheme to reduce the resource cost. Based on the content-match engine and context-aware features, the proposed push notification service can support both of group push notification and bi-directional push notification. Tests are conducted for performance evaluation.

Virtualization has become an important technology that is used across many platforms, particularly servers, to increase utilization, multi-tenancy and security. Virtualization introduces additional overhead that often relates to memory management, interrupt handling and hypervisor mode switching. Among those, memory management and translation lookaside buffer (TLB) management have been shown to have a significant impact on the performance of systems. Two principal mechanisms for TLB management exist in today's systems, namely software and hardware managed TLBs. In this paper, we analyze and quantify the overhead of a pure software virtualization that is implemented over a software managed TLB. We then describe our design of hardware extensions to support virtualization in systems with software managed TLBs to remove the most dominant overheads. These extensions were implemented in the Power embedded A2 core, which is used in the PowerEN and in the Blue Gene/Q processors. They were used to implement a KVM port. We evaluate each of these hardware extensions to determine their overall contributions to performance and efficiency. Collectively these extensions demonstrate an average improvement of 232% over a pure software implementation.

The transistor density of microprocessors continues to increase as technology scales. Microprocessors designers have taken advantage of the increased transistors by integrating a significant number of cores onto a single die. However, a large number of cores are met with diminishing returns due to software and hardware scalability issues and hence designers have started integrating on-chip special-purpose logic units (i.e., accelerators) that were previously available as PCI-attached units. It is anticipated that more accelerators will be integrated on-chip due to the increasing abundance of transistors and the fact that not all logic can be powered at all times due to power budget limits. Thus, on-chip accelerator architectures deserve more attention from the research community. There is a wide spectrum of research opportunities for design and optimization of accelerators. This paper attempts to bring out some insights by studying the data access streams of on-chip accelerators that hopefully foster some future research in this area. Specifically, this paper uses a few simple case studies to show some of the common characteristics of the data streams introduced by on-chip accelerators, discusses challenges and opportunities in exploiting these characteristics to optimize the power and performance of accelerators, and then analyzes the effectiveness of some simple optimizing extensions proposed.

We present the design, optimization and analysis of a highly flexible and efficient multi-leg stock trading system. Automated electronic multi-leg trading allows atomic processing of consolidated orders such as "Buy 200 shares of IBM and sell 100 shares of HPQ". While the expressive power of multi-leg trading brings significant value to investors, it also poses major challenges to stock exchange architecture design, due to additional complexities introduced in performance, tradability, and fairness. Performance can be significantly worse due to the need to coordinate transactions among multiple stocks at once. This paper studies the performance of multi-leg trading under different fairness constraints and variability in order price and order quantity. We identify the major performance bottlenecks when using traditional atomic commitment protocols such as 2-Phase Commit (2PC), and propose a new look-ahead algorithm to maximize transaction concurrency and minimize performance degradation. We have implemented a base-line 2PC prototype and a look-ahead optimized prototype on IBM z10 z Series e Server mainframes. Our experimental results show that the look-ahead optimization can improve throughput by 58% and reduce latency by 30%.

This paper proposes an effective algorithm for spatial error concealment with accurate edge detection and partitioning interpolation. Firstly, a new method is used for detecting possible edge pixels and their matching pixels around the lost block. Then, the true edge lines can be determined, with which the lost block is partitioned. Finally, based on the partition result, each lost pixel can be interpolated with correct reference pixels, which are in the same region with the lost pixel. Experimental results show that the proposed spatial error concealment method is obviously superior to the previous methods for different sequences by up to 4.04 dB.

Although dynamic binary translators (DBT) are gaining popularity in the modern virtual execution environments (VEE), the requirement of DBTs' processing and memory resources has seriously hampered the performance of host platform. In this paper, we propose a distributed DBT system--DistriBit for resource-limited thin clients to overcome these challenges.

This paper discusses the network covert timing channel. This channel modulates network packet's time properties to transfer information secretly. Much work has been done in inventing and utilizing network covert timing channels, however, there is not so much work in other areas such as detecting and handling covert channels. Covert channel detection is difficult, what's more, it often needs human analysis to confirm whether a suspect is a real covert channel. However, we figured out that we can try to control covert channels without knowing whether there are covert channels in using, which means our approach does not rely on the detection of covert channels at all. A network traffic controller is proposed to undermine the network covert timing channel communication mechanism. We will show how our method works and why our traffic control strategy is especially efficient to handle network covert timing channels.

As one of the critical data path functions for many emerging networking applications, packet classification is gaining more and more concerns nowadays. It is commonly believed that conventional software-based classification algorithms are much more time-consuming than hardware-based solutions, i.e., the costly and power consuming TCAM-based mechanism, and incompetent for future high-end applications. In this paper, we propose an efficient optimization framework which can be applied to "gear up" most exiting software-based packet classification algorithms. Under this framework, the large rule set is pre-partitioned into several small subsets, according to some heuristics and dedicated methods. Then the conventional classification process can be significantly simplified and results in a distinct performance improvement by converging the classification power on only a small portion of the rule set. According to the results of our experiment, in which the framework is applied to one of the best algorithms EGT-PC [2], the memory accesses can even be reduced by up to 70%. This provides a much lower cost and more power-efficient alternative to TCAM-based solutions. Another advantage is that the framework requires no change to the hardware environment and little system cost overhead, making it especially suitable for the modern network processor based network solutions.

Most existing packet classification algorithms are decision-tree-based, which are easy-implemented and have reasonably good performance. However, due to rule duplication caused by rule overlapping, these solutions always occupy large and indeterministic-sized memory space, making them only be implemented with mass-but-slow SDRAMs and incompetent for future high-end applications. In this paper, by proposing a novel pre-partitioning scheme for the rule set, we eliminate the overlap among the rules and, therefore, avoid rule duplications. We then further develop a parallel packet classification scheme and a quarter-cut decision tree algorithm. The large rule set is distributedly stored in multiple search engines and the classification can be performed in parallel, independently. Much smaller and more deterministic memory requirement is achieved, as well as O(Log(N)) processing delays. Experimental results show that for a policy table with over 1500 rules, only 36 Kbytes memory is required, and averagely 12 memory accesses are needed for each classification.