Large Multimodal Models (LMMs) have shown promising in-context learning (ICL) capabilities, but scaling to many-shot settings remains difficult due to limited context length and high inference cost. To address these challenges, task-vector-based methods have been explored by inserting compact representations of many-shot in-context demonstrations into model activations. However, existing task-vector-based methods either overlook the importance of where to insert task vectors or struggle to determine suitable values for each location. To this end, we propose a novel Sensitivity-aware Task Vector insertion framework (STV) to figure out where and what to insert. Our key insight is that activation deltas across query-context pairs exhibit consistent structural patterns, providing a reliable cue for insertion. Based on the identified sensitive-aware locations, we construct a pre-clustered activation bank for each location by clustering the activation values, and then apply reinforcement learning to choose the most suitable one to insert. We evaluate STV across a range of multimodal models (e.g., Qwen-VL, Idefics-2) and tasks (e.g., VizWiz, OK-VQA), demonstrating its effectiveness and showing consistent improvements over previous task-vector-based methods with strong generalization.

Reinforcement learning for LLM agents is typically conducted on a static data distribution, which fails to adapt to the agent's evolving behavior and leads to poor coverage of complex environment interactions.To address these challenges, we propose CoEvolve, an agentdata mutual evolution framework that enables LLM agents to improve through closed-loop, interaction-driven training.Specifically, CoEvolve extracts feedback signals such as forgetting and uncertainty from rollout trajectories to identify failure-prone interaction patterns, and utilizes them to guide LLM-based task synthesis.The synthesized tasks are validated through environment interaction and utilized to update the data distribution, enabling joint adaptation of the agent and its data.Extensive experiments on AppWorld and BFCL across Qwen2.5-7B,Qwen3-4B, and Qwen3-30B-A3B demonstrate consistent and significant improvements over strong base models, yielding absolute gains of 19.43%, 15.58%, and 18.14%, respectively.

Object detection via deep neural networks has undergone considerable advancements in recent years. Yet, the detection of smaller objects, specifically those with a few pixels (i.e.,2pixels), is still challenging compared with large objects (i.e., > 962pixels). Existing methods commonly apply high-resolution features or complex super-resolution strategies based on the two-stage Faster Region Convolutional Neural Network (RCNN). They sequentially apply localization and classification stages after a shared feature map extracted by one single backbone network. However, these methods cause low detection accuracy of small objects, high computational overhead, and waste of hardware resources. In this paper, we develop a high-accuracy and real-time small object detection system with negligible computational overhead and low hardware idleness. At the software level, we propose a two-stage Coarse-to-Fine Decoupling RCNN (CFD RCNN) with three techniques: (1) The shared backbone decoupling for localization and classification to achieve high accuracy for both tasks; (2) The training method using backbone feature upsampling for localization with low computational overhead; (3) The object cropping strategy from the original high-resolution image for high-accuracy classification. At the hardware level, we propose a virtualized FPGA accelerator with the Dynamic Resource Allocation (DRA) strategy. The DRA strategy reallocates the hardware resources, considering the workload and resource preference of each stage in CFD RCNN to reduce hardware idleness. Extensive experiments on the TT100K and GTSDB datasets using Xilinx ZCU102 FPGA show that the proposed small object detection system can achieve 2.9% improvement in mean average precision (mAP) compared with state-of-the-art (SOTA) algorithms and raised the throughput from 18.9 FPS to > 26.0 FPS (~1.37×) compared with existing accelerators.

Neural architecture search (NAS) depends heavily on an efficient and accurate performance estimator. To speed up the evaluation process, recent advances, like differentiable architecture search (DARTS) and One-Shot approaches, instead of training every model from scratch, train a weight-sharing super-network to reuse parameters among different candidates, in which all child models can be efficiently evaluated. Though these methods significantly boost search efficiency, they inherently suffer from inaccurate and unstable performance estimation. To this end, we propose a general and effective framework for powering weight-sharing NAS, namely, PWSNAS, by shrinking search space automatically, i.e., candidate operators will be discarded if they are less important. With the strategy, our approach can provide a promising search space of a smaller size by progressively simplifying the original search space, which can reduce difficulties for existing NAS methods to find superior architectures. In particular, we present two strategies to guide the shrinking process: detect redundant operators with a new angle-based metric and decrease the degree of weight sharing of a super-network by increasing parameters, which differentiates PWSNAS from existing shrinking methods. Comprehensive analysis experiments on NASBench-201 verify the superiority of our proposed metric over existing accuracy-based and magnitude-based metrics. PWSNAS can easily apply to the state-of-the-art NAS methods, e.g., single path one-shot neural architecture search (SPOS), FairNAS, ProxylessNAS, DARTS, and progressive DARTS (PDARTS). We evaluate PWSNAS and demonstrate consistent performance gains over baseline methods.

Obstacle avoidance is essential for mobile robot when applying to dynamic scenarios. Existing methods that based on deep reinforcement learning (DRL) use the position information as the environment states and neural network input to train the robot in a low efficiency manner, because the position information is unable to indicate obstacle’s motion trend. To address this problem, this paper proposes a new motion-prediction-based obstacle avoidance method for mobile robots based on DRL. The position information of dynamic obstacles in time domain is used to construct a motion trend vector, and together with other motion state factors to form the robot motion guidance matrix, which effectively expresses the motion change trend of dynamic obstacles in a period that provides more valuable information for the robot to choose avoidance action. The experimental results show that the safety of avoiding dynamic obstacles is effectively improved.

In intelligent manufacturing scenarios, efficient path planning methods are of great significance to the safe and stable operation of multiple automatic guided vehicles (AGVs). This paper contributes a path planning method of multi-AGV considering waiting time optimization, which improves the operation efficiency of multi-AGV. Firstly, aiming at the problems of large granularity and low utilization of spatial resources in traditional topology map planning, an augmented topology map is proposed to improve the utilization rate of geospatial information. Based on this map structure, an improved DFS (depth first search) algorithm is presented to obtain multiple alternative paths during a single search. Moreover, based on the selected path, a corresponding model is established to optimize the waiting time of multi-AGV motion planning scheme, which is solved by genetic algorithm. In addition, the simulations on the multi-AGV platform validate the efficiency of the suggested path planning method.

As an external manifestation of human emotions, expression recognition plays an important role in human-computer interaction. Although existing expression recognition methods performs perfectly on constrained frontal faces, there are still many challenges in expression recognition in natural scenes due to different unrestricted conditions. Expression classification belongs to a pattern recognition problem where intra-class distance is greater than the inter-class distance, which leads to severe over-fitting when using neural networks for expression recognition. This paper proposes a novel net-work structure called Dimension Reduction Network which can effectively reduce generalization error. By adding a data dimension reduction module before the general classification network, a lot of redundant information is filtered, and only useful information is left. This can reduce the interference by irrelevant information when performing classification tasks and reduce generalization error. The proposed method does not require any modification to the classification network, only a small dimension reduction module needs to be added in front of the classification network. However, it can effectively reduce generalization error. We designed big and tiny versions of Dimension Reduction Network, both exceeds our baseline on AffectNet data set. The big version of our proposed method surpassed the state-of-the-art methods by more than 1.2% on AffectNet data set. Our code will open source3when the paper is accepted.

Deep neural networks (DNNs) have achieved great success in a wide range of computer vision areas, but the applications to mobile devices is limited due to their high storage and computational cost. Much efforts have been devoted to compress DNNs. In this paper, we propose a simple yet effective method for deep networks compression, named Cluster Regularized Quantization (CRQ), which can reduce the presentation precision of a full-precision model to ternary values without significant accuracy drop. In particular, the proposed method aims at reducing the quantization error by introducing a cluster regularization term, which is imposed on the full-precision weights to enable them naturally concentrate around the target values. Through explicitly regularizing the weights during the re-training stage, the full-precision model can achieve the smooth transition to the low-bit one. Comprehensive experiments on benchmark datasets demonstrate the effectiveness of the proposed method.

Channel pruning, which seeks to reduce the model size by removing redundant channels, is a popular solution for deep networks compression. Existing channel pruning methods usually conduct layer-wise channel selection by directly minimizing the reconstruction error of feature maps between the baseline model and the pruned one. However, they ignore the feature and semantic distributions within feature maps and real contribution of channels to the overall performance. In this paper, we propose a new channel pruning method by explicitly using both intermediate outputs of the baseline model and the classification loss of the pruned model to supervise layer-wise channel selection. Particularly, we introduce an additional loss to encode the differences in the feature and semantic distributions within feature maps between the baseline model and the pruned one. By considering the reconstruction error, the additional loss and the classification loss at the same time, our approach can significantly improve the performance of the pruned model. Comprehensive experiments on benchmark datasets demonstrate the effectiveness of the proposed method.

In recent years, Convolutional Neural Network (CNN) has been widely applied in computer vision tasks and has achieved significant improvement in image object detection. The CNN methods consume more computation as well as storage, so GPU is introduced for real-time object detection. However, due to the high power consumption of GPU, it is difficult to adopt GPU in mobile applications like automatic driving. The previous work proposes some optimizing techniques to lower the power consumption of object detection on mobile GPU or FPGA. In the first Low-Power Image Recognition Challenge (LPIRC), our system achieved the best result with mAP/Energy on mobile GPU platforms. We further research the acceleration of detection algorithms and implement two more systems for real-time detection on FPGA with higher energy efficiency. In this paper, we will introduce the object detection algorithms and summarize the optimizing techniques in three of our previous energy efficient detection systems on different hardware platforms for object detection.

In recent years, Convolutional Neural Networks (CNNs) have been widely applied in computer vision and have achieved significant improvements in object detection tasks. Although there are many optimizing methods to speed up CNN-based detection algorithms, it is still difficult to deploy detection algorithms on real-time low-power systems. Field-Programmable Gate Array (FPGA) has been widely explored as a platform for accelerating CNN due to its promising performance, high energy efficiency, and flexibility. Previous works show that the energy consumption of CNN accelerators is dominated by the memory access. By fusing multiple layers in CNN, the intermediate data transfer can be reduced. However, previous accelerators with the cross-layer scheduling are designed for a particular CNN model. In addition to the memory access optimization, the Winograd algorithm can greatly improve the computational performance of convolution. In this article, to improve the flexibility of hardware, we design an instruction-driven CNN accelerator, supporting the Winograd algorithm and the cross-layer scheduling, for object detection. We modify the loop unrolling order of CNN, so that we can schedule a CNN across different layers with instructions and eliminate the intermediate data transfer. We propose a hardware architecture to support the instructions with Winograd computation units and reach the state-of-the-art energy efficiency. To deploy image detection algorithms onto the proposed accelerator with fixed-point computation units, we adopt the fixed-point fine-tune method, which can guarantee the accuracy of the detection algorithms. We evaluate our accelerator and scheduling policy on the Xilinx KU115 FPGA platform. The intermediate data transfer can be reduced by more than 90% on the VGG-D CNN model with the cross-layer strategy. Thus, the performance of our hardware accelerator reaches 1700GOP/s on the classification model VGG-D. We also implement a framework for object detection algorithms, which achieves 2.3× and 50× in energy efficiency compared with GPU and CPU, respectively. Compared with floating-point algorithms, the accuracy of the fixed-point detection algorithms only drops by less than 1%.

In recent years, Convolutional Neural Network (CNN) has been widely applied in computer vision tasks. FPGAs have been widely explored to accelerate CNNs due to its high performance, high energy efficiency, and flexibility. By fusing multiple layers in CNN, the intermediate data transfer can be reduced. With a faster algorithm using Winograd transformation, the computation of convolution can be further accelerated. However, previous accelerators with cross-layer or Winograd algorithm are designed for a particular CNN model. The FPGA should be reprogrammed when running another CNN model on the hardware. In this work, we design an instruction driven CNN accelerator supporting Winograd algorithm and cross-layer scheduling. We firstly modify the cross-layer loop unrolling order to extract basic operations as instructions, and then improve the on-chip memory architecture for higher computation units utilization rate in Winograd. We evaluate the hardware architecture and scheduling policy on Xilinx Virtex-7 690t FPGA platform. As a case study, the intermediate data transfer can be reduced by over 90% on VGG-D CNN model with cross-layer policy. The performance of our hardware accelerator reaches 1500 GOP/s. Experimental results show that our design achieves a 7 χ speed-up than previous cross-layer FPGA accelerator on the same platform. The performance can be further improved by 78% if larger Winograd transformation sizes are used.

In this paper, we propose PGIS, a parallelism and garbage collection aware I/O Scheduler, which identifies the hot data based on trace characteristics to exploit the channel level internal parallelism of flash-based storage systems. PGIS not only fully exploits abundant channel resource in the SSD, but also it introduces a hot data identification mechanism to reduce the garbage collection overhead. By dispatching hot read data to different channel, the channel level internal parallelism is fully exploited. By dispatching hot write data to the same physical block, the garbage collection overhead has been alleviated. The experiment results show that compared with existing I/O schedulers, PGIS improves the response time and garbage collection performance significantly. Consequently, PGIS reduces the garbage collection overhead up to 30.9%, while exploiting channel level internal parallelism.

Genetic risk prediction is an important goal in human genetics research and precision medicine. Accurate prediction models will have great impacts on both disease prevention and early treatment strategies. Despite the identification of thousands of disease-associated genetic variants through genome wide association studies (GWAS), genetic risk prediction accuracy remains moderate for most diseases, which is largely due to the challenges in both identifying all the functionally relevant variants and accurately estimating their effect sizes in the presence of linkage disequilibrium. In this paper, we introduce AnnoPred, a principled framework that leverages diverse types of genomic and epigenomic functional annotations in genetic risk prediction for complex diseases. AnnoPred is trained using GWAS summary statistics in a Bayesian framework in which we explicitly model various functional annotations and allow for linkage disequilibrium estimated from reference genotype data. Compared with state-of-the-art risk prediction methods, AnnoPred achieves consistently improved prediction accuracy in both extensive simulations and real data.

MOTIVATION: Genome-wide association study (GWAS) has been a great success in the past decade. However, significant challenges still remain in both identifying new risk loci and interpreting results. Bonferroni-corrected significance level is known to be conservative, leading to insufficient statistical power when the effect size is moderate at risk locus. Complex structure of linkage disequilibrium also makes it challenging to separate causal variants from nonfunctional ones in large haplotype blocks. Under such circumstances, a computational approach that may increase signal replication rate and identify potential functional sites among correlated markers is urgently needed. RESULTS: We describe GenoWAP, a GWAS signal prioritization method that integrates genomic functional annotation and GWAS test statistics. The effectiveness of GenoWAP is demonstrated through its applications to Crohn's disease and schizophrenia using the largest studies available, where highly ranked loci show substantially stronger signals in the whole dataset after prioritization based on a subset of samples. At the single nucleotide polymorphism (SNP) level, top ranked SNPs after prioritization have both higher replication rates and consistently stronger enrichment of eQTLs. Within each risk locus, GenoWAP may be able to distinguish functional sites from groups of correlated SNPs. AVAILABILITY AND IMPLEMENTATION: GenoWAP is freely available on the web at http://genocanyon.med.yale.edu/GenoWAP.

Solid State Drives (SSDs) have become very popular recently due to their high performance and other benefits such as shock-resistance. However, SSDs pose some unique and serious challenges to I/O and file system designers because of flash memory's unique properties, such as out-of-place update, wearing-out, and highly asymmetric performance for read, write and erase operations. Most SSDs employ a log-structured block-based FlashTranslation-Layer (FTL) to solve the out-of-place update problem. The performance of FTLs is often highly sensitive to access patterns, especially the write access patterns. For example, sequential write requests see lower overhead than random writes. Moreover, sequential write requests that are not aligned to the flash page boundary may cause extra write and garbage collection operations, increasing overhead and wear-out. In this paper, we present a novel write buffer design based on sophisticated, fine-grain write access pattern analysis. Our scheme identifies access patterns in a per-process per-stream granularity in the OS buffer cache. These patterns are then used to guide the write buffer to improve the write performance of SSDs that employ a log-structured block-based FTL. Simulation results show that our solutions can improve write performance by up to 38%. Moreover, the schemes reduce SSD erase cycles by up to 56%, which is directly translated to a major improvement on the life-span of SSDs.

Most supercomputers nowadays are based on large clusters, which call for sophisticated, scalable, and decentralized metadata processing techniques. From the perspective of maximizing metadata throughput, an ideal metadata distribution policy should automatically balance the namespace locality and even distribution without manual intervention. None of existing metadata distribution schemes is designed to make such a balance. We propose a novel metadata distribution policy, Dynamic Dir-Grain (DDG), which seeks to balance the requirements of keeping namespace locality and even distribution of the load by dynamic partitioning of the namespace into size-adjustable hierarchical units. Extensive simulation and measurement results show that DDG policies with a proper granularity significantly outperform traditional techniques such as the Random policy and the Subtree policy by 40 percent to 62 times. In addition, from the perspective of file system reliability, metadata consistency is an equally important issue. However, it is complicated by dynamic metadata distribution. Metadata consistency of cross-metadata server operations cannot be solved by traditional metadata journaling on each server. While traditional two-phase commit (2PC) algorithm can be used, it is too costly for distributed file systems. We proposed a consistent metadata processing protocol, S2PC-MP, which combines the two-phase commit algorithm with metadata processing to reduce overheads. Our measurement results show that S2PC-MP not only ensures fast recovery, but also greatly reduces fail-free execution overheads.

Recent research at Intel suggests that chips with hundreds of processor cores are possible in the not-so-distant future. As the number of cores grows, so does the size of the cache systems required to allow them to operate efficiently. Caches have grown to consume a significant percentage of the power utilized by a processor. In this research, we extend the concept of location cache to support chip multiprocessors (CMPs) systems in combination with low-power L2 caches based upon the gated-ground technique. The combination of these two techniques allows for reductions in both dynamic and leakage power consumption. In this paper, we will present an analysis of the power savings provided by utilizing location caches in a CMP system. The performance of the cache system is evaluated by extending the capability of CACTI and Simics using the SPLASH-2 and ALPBench benchmark suites. These simulation results demonstrate that the utilization of location caches in CMP systems is capable of saving a significant amount of power over equivalent CMP systems that lack location caches.

To improve the performance of similarity search and information retrieval is an important research issue in peer-to-peer environment. In this paper, we propose a distributed architecture for enhancing the performance of similarity search in unstructured P2P networks. The key component of the proposed architecture is a distributed, content-based, heuristic feedback mechanism, which allows peers to keep track of recent queries and learn from the assessment of answers to previous queries, so as to self-adaptively route the subsequent query to the most relevant nodes which are responsible for the query. Therefore a high recall rate can be achieved by probing only a small amount of peers. We also propose a distributed automatic query expansion mechanism to improve the quality of query results. Since the architecture is entirely distributed, it scales well with the large sized networks. The experimental results show that our architecture can efficiently solve queries with a relatively small cost.

Data availability is a challenging issue for large-scale cluster file systems built upon thousands of individual storage devices. Replication is a well-known solution used to improve data availability. However, how to efficiently guarantee replicas consistency under concurrent conflict mutations remains a challenge. Moreover, how to quickly recover replica consistency from a storage server crash or storage device failure is also a tough problem. In this paper, we present a replication-based data availability mechanism designed for a large-scale cluster file system prototype named LionFS. Unlike other replicated storage systems that serialize replica updates, LionFS introduces a relaxed consistency model to enable concurrent updating all replicas for a mutation operation, greatly reducing the latency of operations. LionFS ensures replica consistency if applications use file locks to synchronize the concurrent conflict mutations. Another novelty of this mechanism is its light-weight log, which only records failed mutations and imposes no overhead on failure-free execution and low overhead when some storage devices are unavailable. Furthermore, recovery of replica consistency needs not stop the file system services and running applications. Performance evaluation shows that our solution achieves 50–70% higher write performance than serial replica updates. The logging overhead is shown to be low, and the recovery time is proportional to the amount of data written during the failure.

Publish/subscribe model has become a prevalent paradigm for building distributed event delivering systems. Content-based publish/subscribe allows high expresses in subscriptions and thus is more appropriate for content dissemination. However, the scalability has remained a challenge in the design of distributed content-based publish/subscribe systems due to the expensive matching and delivering cost of content-based events. In this paper we propose HyperSub, an infrastructure built on top of distributed hash table (DHT) for efficient content-based data distribution. Based on efficient subscription installation, event publication and event delivery mechanisms, Hyper- Sub can provide a scalable platform to simultaneously support any numbers of pub/sub schemes with different number of attributes. There are three key features in the design of HyperSub: (1) a locality-preserving hashing mechanism which partitions the content space into content zones. Subscriptions and events are mapped to the content zones for efficiently matching; (2) an efficient event delivery algorithm which exploits the embedded trees in the underlying DHT to deliver events to the corresponding subscribers; (3) light-weighted load balancing mechanisms to adjust the load among peers and ensure that no peer is unduly loaded.

The indexing of complex data and similarity search plays an important role in many application areas. Traditional centralized index structure can not scale with the rapid proliferation of data volume. In this paper, we propose a scalable index architecture built on top of distributed hash tables (DHT), to support similarity search in the general metric space. Based on efficient space mapping and query routing mechanisms, our architecture can provide a general platform to support arbitrary number of indexes on different data types. Significantly, it does not need to generate or maintain any search trees. Instead, the embedded trees in the underlying distributed hash tables are exploited to deliver queries. To deal with skewed data distribution, we also provide load-balancing mechanisms to ensure that no node in the system is unduly loaded. The performance of the proposed design is evaluated through simulations with a variety of metrics. The experimental results demonstrate that out approach can efficiently solve similarity query at a low cost.

One hurdle to using peer-to-peer networks as anonymizing networks is churn. Node churn makes anonymous paths fragile and short-lived: failures of a relay node disrupt the path, resulting in message loss and communication failures. To make anonymous routing resilient to node failures, we take two steps: (1) we use a simple yet powerful idea based on message redundancy by erasure coding and path redundancy to mask node failures; (2) we base mix choices of a path on node lifetime prediction and choose stable nodes as relay nodes, thereby prolonging single path durability. We present an allocation of erasure-coded message segments among multiple paths that provides a guideline on how to maximize routing resilience upon different node availabilities in real-world systems. Via detailed simulations, we compare routing resilience of our approach and existing anonymity protocols, showing that our approach greatly improves routing resilience while consuming modest bandwidth.

Leveraging the state-of-the-art information retrieval (IR) algorithms and relevance ranking algorithm, we propose an architecture for efficient search in peer-to-peer networks. The key idea is to employ a distributed, content-based algorithm to arrange the participating nodes into semantic clusters, in which most connections are short-distance, connecting pairs of nodes with similar contents. Nodes in different clusters are connected through long-distance links. Queries are routed to the cluster which has the most relevance contents and flooded to the nodes in the corresponding clusters. The experiment results show that the proposed architecture can achieve high recall rate with small cost of network traffic and query processing load.

The scalability has remained a challenge in the design of distributed publish/subscribe systems. In this paper we propose a novel solution to address this problem in content-based pub/sub systems on top of distributed hash table. The main objective is to ensure an appropriate amount of rendezvous point nodes in the system, as well as maintain an even load distribution among them. An attribute-vector based publish/subscribe scheme and related load balancing mechanisms (ID space partitioning, attribute grouping, dynamic ID space split-merge) are proposed to achieve this goal. The experimental results show that our approaches can achieve a good scalability by efficiently distribute/balance load among an adaptive quantity of rendezvous point nodes, while retaining very small overhead and latency

Due to the spatial-locality property of data caches and the temporal-locality property of instruction caches, significant leakage reduction can be achieved by switching a large number of cache lines into the low-power standby or drowsy mode. It has been shown that 80%-90% of the data cache lines can be maintained in drowsy state without affecting the performance by more than 0.6% (IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 12, no. 2, pp. 167-184, Feb. 2004). However, with the introduction of the drowsy-cache design technique, new fault behaviors appear and more restrictive design rules must be applied to the chip fabrication process. In this paper, we simulate all possible spot defects (SDs) under normal mode and drowsy mode in different resistance regions using HSpice. Six new fault models appear with the introduction of drowsy mode for memory arrays. When we derive a march algorithm for the new fault models of this low-power cache, several simplification rules are utilized to reduce the test complexity. According to these simplification rules, each of these new faults has its equivalent counterpart existent in both data caches and instruction caches. As a result, we develop a march algorithm which can detect all SDs in either data caches or instruction caches. Since some faults occur only in drowsy mode, a built-in self-repair (BISR) scheme is developed. By utilizing BISR, the cache can still work even if some cache lines fail to work in drowsy mode

We propose Ferry, an architecture that extensively yet wisely exploits the underlying distributed hash table (DHT) overlay structure to build an efficient and scalable platform for content-based publish/subscribe (pub/sub) services. Ferry aims to host any and many content-based pubservices: Any pubsub service with a unique scheme can run on top of Ferry, and multiple pub/sub services can coexist on top of Ferry. For each pub/sub service, Ferry does not need to maintain or dynamically generate any dissemination tree. Instead, it exploits the embedded trees in the underlying DHT to deliver events, thereby imposing little overhead. Ferry can support a pub/sub scheme with a large number of event attributes. To deal with a skewed distribution of subscriptions and events, Ferry uses one-hop subscription push and attribute partitioning to balance load

Distributed hash table based peer-to-peer systems are emerging as new paradigms for building large-scale distributed applications, due to their scalability, fault-tolerance and self-organization. However, most existing DHTs are designed for exact-key searching, and the support of multi-dimensional range queries in peer-to-peer networks is still a challenging problem. In this paper, we propose a distributed index architecture called Dak to support range queries on multi-dimensional data. Based on efficient space mapping and query routing mechanisms, Dak can provide a scalable platform to support any number of indexes with different dimensionalities. Significantly, this architecture does not need to generate or maintain any search trees. Instead, it exploits the embedded trees in the underlying distributed hash tables to refine and deliver queries. To deal with skewed data distribution, we also provide load-balancing mechanisms to ensure that no node in the system is unduly loaded

Content delivery network (CDN) plays an important role in today's Web services. More and more content providers use CDNs to lower the server overhead, reduce client perceived latency and decrease network traffic. More research papers have been published in recent years addressing CDN system performance issues. However, most of them are concentrated on server placement policy, content distribution mechanism and request routing algorithm. In this paper, we propose a new CDN architecture to improve system performance by grouping the content servers which are topologically close into server clusters and exploiting the benefits of server cooperation. In our approach, when a server receives a request and cannot fulfil this request, it will forward the request to other nearby servers in this cluster. If there is a cache bit, the data can be fetched immediately. Only in case none of the servers can satisfy this request, the data will be fetched from the original server. Furthermore, with the server coordination, system workload can be well balanced on several servers. We conduct extensive simulations and the results show that our solution achieves significant improvement over the conventional CDN architecture

With the introduction of the drowsy cache design technique (Kim et al., 2004), new fault behaviors appear and more restrictive design rules must be applied to the chip fabrication process. In this research, we simulate all possible spot defects (SDs) under normal mode and drowsy mode in different resistance regions using HSpice. Six new fault models appear with the introduction of drowsy mode for memory arrays. When we derive a march algorithm for the new fault models of this low-power cache, several simplification rules are utilized to reduce the test complexity. As a result, we develop a march algorithm which can detect all SDs in either data caches or instruction caches. Since some faults occur only in drowsy mode, a built-in self-repair (BISR) scheme is developed. By utilizing BISR, the cache can still work even if some cache lines fail to work in drowsy mode

The big challenges facing the search techniques on Gnutella-like peer-to-peer networks are search efficiency and quality of search results. In this paper, leveraging information retrieval (IR) algorithms such as Vector Space Model (VSM) and relevance ranking algorithms, we present GES (Gnutella with Efficient Search) to improve search performance. The key idea is that GES uses a distributed topology adaptation algorithm to organize semantically relevant nodes into same semantic groups by using the notion of node vector. Given a query, GES employs an efficient search protocol to direct the query to the most relevant semantic groups for answers, thereby achieving high recall with probing only a small fraction of nodes. To the best of our knowledge, GES is the first to identify node vector size as an important role in impacting search performance and to show that the node vector size offers a good trade-off between search performance and bandwidth cost. Moreover, GES adopts automatic query expansion and local data clustering to improve search performance. We show that GES is efficient and even outperforms the centralized node clustering system SETS. For example, in the scenario where node capacity is heterogeneous, GES can achieve 73 percent recall when probing only 20 percent nodes, outperforming SETS by about 18 percent.

Leveraging DHTs (distributed hash table), we propose Ferry, an architecture for content-based publish/subscribe services. With its novel design in subscription installation, subscription management and event delivery algorithm, Ferry can serve as a scalable platform to host any and many content-based publish/subscribe services: any publish/subscribe service with a unique scheme can run on top of Ferry, and many publish/subscribe services can run together on top of Ferry. For each of the publish/subscribe services running on top of Ferry, Ferry does not need to maintain or dynamically generate any dissemination tree. Instead, it exploits the embedded trees in DHTs such as Chord to deliver events, thereby imposing little overhead. By delivering events along DHT links, Ferry has two main advantages: (1) it eliminates the cost in construction and maintenance of the dissemination trees; (2) it allows further optimization, i.e., the DHT link maintenance messages could be piggybacked onto the event delivery messages to reduce the maintenance cost which is inherent and nontrivial in DHTs. Moreover, Ferry can support a publish/subscribe scheme with a very large number of event attributes.

This paper presents the design and implementation of a different approach called RAPID-cache (redundant, asymmetrically parallel and inexpensive disk cache) for highly-reliable systems. In order to validate whether RAPID-cache can really achieve satisfactory performance in real systems, and to see if RAPID-cache can also improve the performance/reliability of the OS buffer cache, we have designed and implemented RAPID-cache for the Linux system. Measurement results using realistic benchmarks show very promising results. Using only 3.25 MB of NVRAM, a RAPID-Cache outperforms normal NVRAM caches of 64 MB or even 128 MB. RAPID-Cache also provide better reliability than normal NVRAM caches.

Most of structured P2P overlays exploit distributed Hash table (DHT) to achieve an administration-free, fault tolerant overlay network and guarantee to deliver a message to the destination within O(log N) hops. While elegant from a theoretical perspective, those systems face difficulties under an open environment. Not only frequently joining and leaving of nodes generate enormous maintenance overhead, but also various behaviors and resources of peers compromise overlay performance and security. Instead of building P2P overlay from the theoretical view, this paper builds the overlay network from the perspective of the physical network. By combining different network topology-aware techniques, a distinctive overlay network that closely matches the Internet topology is constructed. The P2P system based on this structure is not only highly efficient for routing, but also keeps maintenance overhead very low even under highly dynamic environment. Moreover, the system provides a new aspect to solve many security issues.

Many solutions have been proposed to tackle the load balancing issue in DHT-based P2P systems. However, all these solutions either ignore the heterogeneity nature of the system, or reassign loads among nodes without considering proximity relationships, or both. In this paper, we present an efficient, proximity-aware load balancing scheme by using the concept of virtual servers. To the best of our knowledge, this is the first work to use proximity information in load balancing. In particular, our main contributions are: 1) relying on a self-organized, fully distributed k-ary tree structure constructed on top of a DHT, load balance is achieved by aligning those two skews in load distribution and node capacity inherent in P2P systems - that is, have higher capacity nodes carry more loads; 2) proximity information is used to guide virtual server reassignments such that virtual servers are reassigned and transferred between physically close heavily loaded nodes and lightly loaded nodes, thereby minimizing the load movement cost and allowing load balancing to perform efficiently; and 3) our simulations show that our proximity-aware load balancing scheme reduces the load movement cost by 11-65 percent for all the combinations of two representative network topologies, two node capacity profiles, and two load distributions of virtual servers. Moreover, we achieve virtual server reassignments in O(log N) time.

We present TAP, a novel tunneling approach for anonymity in structured P2P systems. An important feature of TAP is that anonymous tunnels are fault-tolerant to node failures. Relying on P2P routing infrastructure and replication mechanism, the basic idea behind TAP is to decouple anonymous tunnels from "fixed" P2P nodes and form anonymous tunnels from dynamic tunnel hop nodes. The primary motivation of TAP is to strike a balance between functionality and anonymity in dynamic P2P networks. We have implemented the tunneling mechanism in Java on FreePastry 1.3. An analysis of its anonymity and performance was evaluated via detailed simulations.

Summary form only given. Web caching is the most important technique to reduce network bandwidth consumption and minimize user-perceived latency. However, most researches are focused on designing efficient architectures with dedicated proxy servers, the potential advantage of utilizing client cache is not fully exploited. We propose a new solution to improve Web caching performance using local cache on client computers. Our system has several advantages than the dedicated proxy server mechanism. First, a larger virtual cache is generated to cache more documents than a single proxy server. Second, it is scalable, system workloads are distributed all across client computers instead of concentrated on a central server, the "hot spot" and "single point of failure " problems are relieved. Third, by the introduction of the superclients, the effect of the weak clients in a fully decentralized scheme is also alleviated. The simulation results show our system can achieve better system caching performance and scalability than the previous solutions.

Summary form only given. Well-organized proxy caching systems can greatly reduce the user perceived latency and decrease the network bandwidth consumption. In this paper, we propose a new hash based Web caching architecture, Tulip. Tulip extends the locality-based algorithm in UCFS as the basic data grouping scheme in hash based proxy systems, uses it to aggregate Web objects which are likely to be accessed together into object clusters and uses these clusters as the primary access units between memory and disk. The overhead of slow disk I/Os is greatly reduced. It also presents a simple and efficient data duplication scheme. Along with the local caching strategy, Tulip can achieve both fault tolerance and load balance with minimal overhead introduced. Our simulation results show Tulip is scalable and robust, it has better performance than previous approaches.

Summary form only given. Many solutions have been proposed to tackle the problem of load balancing in structured P2P systems. However, all these solutions either ignore the heterogeneity nature of the system, or reassign loads among nodes without considering proximity relationships, or both. We propose an efficient scheme to ensure fair load distribution in such P2P systems by utilizing proximity information. To our knowledge, this is the first work to utilize the proximity information to guide load balancing. In particular, our main contributions are: (1) A self-organized, fully distributed K-nary tree structure is constructed on top of a distributed hash table (DHT) for load balancing information collection/dissemination and load reassignment. (2) Load balancing is achieved by aligning those two skews in load distribution and node capacity inhere in P2P systems - that is, have higher capacity nodes carry more loads. (3) Proximity information is utilized to guide load balancing such that loads are assigned and transferred between physically close heavy nodes and light nodes, thereby minimizing the load transferring overhead and allowing load balancing to perform more efficiently.

This work presents a comprehensive performance study on Internet Service Provider (ISP) mail server, which plays an important role in Internet-based distributed computing. By feeding the SPECmail2001 benchmark into an ISP mail server testbed set up by a commercial mail server software system - MDaemon 5.0.1, we study both networking and I/O performance by varying its user population from 200 to 10,000. The benchmark utilities are adopted for networking analysis while file system traces are collected for I/O measurement. Based on the benchmark study and offline trace analysis, we arrive at several important conclusions for ISP mail servers and give corresponding technical suggestions to improve the performance. First we observed that, in SMTP and POP sessions, the initial network connection setup step usually takes a very long time. Second, I/O latencies typically contribute to 40-55% of the total data transfer time in e-mail requests, especially in a server with large user population support. Third, A group of e-mail messages will easily make a remote recipient server become overloaded.

While set-associative caches incur fewer misses than direct-mapped caches, they typically have slower hit times and higher power consumption, when multiple tag and data banks are probed in parallel. This paper presents the location cache structure which significantly reduces the power consumption for large set-associative caches. We propose to use a small cache, called location cache to store the location of future cache references. If there is a hit in the location cache, the supported cache is accessed as a direct-mapped cache. Otherwise, the supported cache is referenced as a conventional set-associative cache.The worst case access latency of the location cache system is the same as that of a conventional cache. The location cache is virtually indexed so that operations on it can be performed in parallel with the TLB address translation. These advantages make it ideal for L2 cache systems where traditional way-predication strategies perform poorly.We used the CACTI cache model to evaluate the power con-sumption and access latency of proposed cache architecture. Simplescalar CPU simulator was used to produce final results. It is shown that the proposed location cache architecture is power-efficient. In the simulated cache configurations, up-to 47% of cache accessing energy and 25% of average cache access latency can be reduced.

Multimedia applications are fast becoming one of the dominating workloads for modern computer systems. Since these applications normally have large data sets and seem to have little data-reuse, a commonly held belief is that they have poor memory behavior compared to traditional programs and that current cache architectures cannot handle them well. It is therefore important to quantitatively characterize the memory behavior of these applications in order to provide insights for future design and research of memory systems. However, very few results on this topic have been published. We analyze the cache behavior of a group of representative multimedia applications. These programs include a subset of the popular MediaBench suite and several large multimedia programs running on the Linux, Windows 98, and TruUnix operating systems. We performed extensive measurement and trace-driven simulation experiments. We then compared the memory utilization of these programs to that of the SPEC2000 applications. We found that multimedia applications actually have better cache behavior than SPEC2000 programs. The following three factors contribute to the high cache hit rates of multimedia applications: First, most multimedia applications apply block-partitioning algorithms to the input data and work on small blocks of data that easily fit in the cache. Second, within these blocks, there is significant data reuse as well as spatial locality. Third, a large number of references generated by multimedia applications are to their internal data structures, which are relatively small and can easily fit in reasonably sized caches.

Routing algorithm has great influence on system overall performance in peer-to-peer (P2P) applications. In current DHT based routing algorithms, routing tasks are distributed across all system peers. However, a routing hop could happen between two widely separated peers with high network link latency which greatly increases system routing overheads. We propose a new P2P routing algorithm - HIERAS to relieve this problem, it keeps scalability property of current DHT algorithms and improves system routing performance by the introduction of hierarchical structure. In HIERAS, we create several lower level P2P rings besides the highest level P2P ring. A P2P ring is a subset of the overall P2P overlay network. We create P2P rings in such a strategy that the average link latency between two peers in lower level rings is much smaller than higher level rings. Routing tasks are first executed in lower level rings before they go up to higher level rings, a large portion of routing hops previously executed in the global P2P ring are now replaced by hops in lower level rings, thus routing overheads can be reduced. The simulation results show HIERAS routing algorithm can significantly improve P2P system routing performance

New demands brought by the continuing growth of the Internet will be met in part by more effective and comprehensive use of caching. We propose to exploit client browser caches in cooperative proxy caching by constructing the client caches within each organization (e.g., corporate networks) as a large peer-to-peer client cache. Via trace-driven simulations we evaluate the potential performance benefit of cooperative proxy caching with/without exploiting client caches. We show that exploiting client caches in cooperative proxy caching can significantly improve performance, particularly when the size of individual proxy caches is limited compared to the universe of Web objects. We further devise a cooperative hierarchical greedy-dual replacement algorithm (Hier-GD), which not only provides some cache coordination but also utilizes client caches. Through Hier-GD, we explore the design issues of how to exploit client caches in cooperative proxy caching to build large Web caches. We show that Hier-GD is technically practical and can potentially improve the performance of cooperative proxy caching by utilizing client caches

Peer-to-peer (P2P) system has become one of the hottest research topics, its excellent characteristics of fully decentralized control and self-organizing make it attractive for some particular applications. However, it faces more technical problems than client/server architecture. In this paper, we propose SBARC,a new P2P file sharing system which takes into account the tremendous resource difference among peers to improve system performance. SBARC divides the peers into supernodes and ordinary nodes and most workloads are taken into supernodes. The main contributions of SBARC are (1). A supernode based routing algorithm which can reduce the average routing latency; (2) Routing information are cached to further reduce the routing cost; (3). A coordinate file caching scheme which achieves efficient utilization of free storage space. Our simulation results show SBARC routing and caching schemes can achieve better performance than previous approaches.

DHT based peer-to-peer (P2P) algorithms are very promising for their efficient routing performance. However, most commercial P2P systems do not adapt DHT algorithms and still use central facilities or broadcasting based routing mechanisms. One reason impeding the DHT algorithm popularity is the routing information maintenance overhead in DHT algorithms; it generates considerable network traffic and increases P2P system complexity, especially in a highly dynamic environment. We discuss its effects on DHT routing performance and propose our solution to reduce this overhead.

Structured peer-to-peer (P2P) systems address the load balancing issue in a rather naive way, by simply resorting to the uniformity of the hash function utilized to generate object IDs. Such a random choice of object IDs could result in O(log N) load imbalance. We propose an efficient, proximity-aware load balancing algorithm for such systems. To our knowledge, this is the first work to utilize the proximity information to guide load balancing. In particular, our main contributions are: (1) A self-organized, fully distributed K-nary tree structure is constructed on top of a DHT for load balancing information collection/dissemination and load reassignment. (2) Load balancing is achieved by aligning those two skews in both load distribution and node capacity inhere in P2P systems - that is, have higher capacity nodes carry more loads. (3) Proximity information is utilized to guide load balancing such that virtual servers are assigned and transferred between physically close heavy nodes and light nodes, thereby minimizing the load transferring overhead and making load balancing fast and efficient.

We present an architecture for a peer-to-peer (P2P) file system which supports semantics-based access. Central to this work is to provide semantic indexing and retrieval capabilities. Our semantic indexing and locating approach is based on distributed hash tables (DHTs) where the indices of semantically close files are clustered to the same peers with high probability (nearly 100%) by the use of locality sensitive hash functions. A query for finding semantically close files can be answered by consulting only a small number of peer nodes which are most responsible for such a query, instead of by query flooding. Our approach only adds index information to peer nodes, thus imposing only a small storage overhead. We constitute an initial step to integrate semantics-based access mechanisms into a P2P file system.

This paper presents a strong security scheme for network-attached storage (NAS) that is based on capability and uses a key distribution scheme to keep network-attached storage from performing key management. Our system uses strong cryptography to protect data from spoofing, tampering, eavesdropping and replay attacks, and it also guarantees that the data stored on the storage is copy-resistant. In spite of this level of security, our system does not impose much performance penalty. Our experimental results shows that, using a relatively inexpensive CPU in the storage device, there are little performance penalty for random disk accesses and about 9-25% performance degradation for large sequential disk accesses (/spl ges/4 KB).

We present a novel reordering write buffer which improves the performance of log-structured file systems (LFS). While UFS has a good write performance, high garbage-collection overhead degrades its performance under high disk space utilization. Previous research concentrated on how to improve the efficiency of the garbage collector after data is written to disk. We propose a new method that reduces the amount of work the garbage collector would do before data reaches disk. By classifying active and inactive data in memory into different segment buffers and then writing them to different disk segments, we force the disk segments to form a bimodal distribution. Most data blocks in active segments are quickly invalidated, while inactive segments remain mostly intact. Simulation results based on a wide range of both real-world and synthetic traces show that our method significantly reduces the garbage collection overhead, slashing the overall write cost of LFS by up to 53 percent, improving the write performance of LFS by up to 26 percent, and the overall system performance by up to 21 percent.

Via detailed file system and disk system simulation, we examine the impact of disk built-in caches on the system performance. Our results indicate that the current trend of using large built-in caches is unnecessary and a waste of money and power for most users. Disk manufacturers could use much smaller built-in caches to reduce the cost as well as power-consumption, without affecting performance.

Web proxy caching servers play a key role in today's Web infrastructure. Previous studies have shown that disk I/O is one of the major performance bottlenecks of proxy servers. Most conventional file systems do not work well for proxy server workloads and have high overheads. This paper presents a novel, User-space, Customized File System, called UCFS, that can drastically improve the I/O performance of proxy servers. UCFS is a user-level software component of a proxy server which manages data on a raw disk or disk partition. Since the entire system runs in the user space, it is easy and inexpensive to implement. It also has good portability and maintainability. UCFS uses efficient in-memory meta-data tables to eliminate almost all I/O overhead of meta-data searches and updates. It also includes a novel file system called Cluster-structured File System (CFS). Similarly to the Log-structured File Systems (LFS), CFS uses large disk transfers to significantly improve disk write performance. However, CFS can also markedly improve file read operations and it does not generate garbage. Comprehensive simulation experiments using five representative real-world traces show that UCFS can significantly improve proxy server performance. For example, UCFS achieves 8-19 times better I/O performance than the state-of-the-art SQUID server running on a Unix Fast File System (FFS), 4-7.5 times better than SQUID on asynchronous FFS, and 3-9 times better than the improved SQUIDML.

Modern high performance disk systems make extensive use of nonvolatile RAM (NVRAM) write caches. A single-copy NVRAM cache creates a single point of failure while a dual-copy NVRAM cache is very expensive because of the high cost of NVRAM. This paper presents a new cache architecture called RAPID-Cache for Redundant, Asymmetrically Parallel, and Inexpensive Disk Cache. A typical RAPID-Cache consists of two redundant write buffers on top of a disk system. One of the buffers is a primary cache made of RAM or NVRAM and the other is a backup cache containing a two-level hierarchy: a small NVRAM buffer on top of a log disk. The small NVRAM buffer combines small write data and writes them into the log disk in large sizes. By exploiting the locality property of I/O accesses and taking advantage of well-known Log-structured File Systems, the backup cache has nearly equivalent write performance as the primary RAM cache. The read performance of the backup cache is not as critical because normal read operations are performed through the primary RAM cache and reads from the backup cache happen only during error recovery periods. The RAPID-Cache presents an asymmetric architecture with a fast-write-fast-read RAM being a primary cache and a fast-write-slow-read NVRAM-disk hierarchy being a backup cache. The asymmetrically parallel architecture and an algorithm that separates actively accessed data from inactive data in the cache virtually eliminate the garbage collection overhead, which are the major problems associated with previous solutions such as Log-structured File Systems and Disk Caching Disk. The asymmetric cache allows cost-effective designs for very large write caches for high-end parallel disk systems that would otherwise have to use dual-copy, costly NVRAM caches. It also makes it possible to implement reliable write caching for low-end disk I/O systems since the RAPID-Cache makes use of inexpensive disks to perform reliable caching. Our analysis and trace-driven simulation results show that the RAPID-Cache has significant reliability/cost advantages over conventional single NVRAM write caches and has great cost advantages over dual-copy NVRAM caches. The RAPID-Cache architecture opens a new dimension for disk system designers to exercise trade-offs among performance, reliability, and cost.

Modern CPUs often use large physically indexed caches that are direct-mapped or have low associativities. Such caches do not interact well with virtual memory systems. An improperly placed physical page will end up in a wrong place in the cache, causing excessive conflicts with other cached pages. Page coloring has been proposed to reduce the conflict misses by carefully placing pages in the physical memory. While page coloring works well for some applications, many factors limit its performance. Page coloring limits the freedom of the page placement system and may increase swapping traffic. In this paper, we propose a novel and simple architecture, called color-indexed, physically tagged caches, which can significantly reduce the conflict misses. With some simple modifications to the TLB (translation look-aside buffer), the new architecture decouples the addresses of the cache from the addresses of the main memory. Since the cache addresses do not depend on the the physical memory addresses anymore, the system can freely place data in any cache page to minimize the conflict misses, without affecting the paging system. Extensive trace-driven simulation results show that our design performs much better than traditional page coloring techniques. The new scheme enables a direct-mapped cache to achieve hit ratios very close to or better than those of a two-way set associative cache. Moreover, the architecture does not increase cache access latency, which is a drawback of set associative caches. The hardware overhead is minimal. We show that our scheme can reduce the cache size by 50 percent without sacrificing performance. A two-way set associative cache that uses this strategy can perform very close to a fully associative cache.

This paper presents a new cache architecture called RAPID-Cache for Redundant, Asymmetrically Parallel, and Inexpensive Disk Cache. A typical RAPID-Cache consists of two redundant write buffers on top of a disk system. One of the buffers is a primary cache made of RAM or NVRAM and the other is a backup cache containing a two level hierarchy: a small NVRAM buffer on top of a log disk. The backup cache has nearly equivalent write performance as the primary RAM cache, while the read performance of the backup cache is not as critical because normal read operations are performed through the primary RAM cache and reads from the backup cache happen only during error recovery periods. The RAPID-Cache presents an asymmetric architecture with a fast-write-fast-read RAM being a primary cache and a fast-write-slow-read NVRAM-disk hierarchy being a backup cache. The asymmetric cache architecture allows cost-effective designs for very large write caches for high-end disk I/O systems that would otherwise have to use dual-copy, costly NVRAM caches. It also makes it possible to implement reliable write caching for low-end disk I/O systems since the RAPID-Cache makes use of inexpensive disks to perform reliable caching. Our analysis and trace-driven simulation results show that the RAPID-Cache has significant reliability/cost advantages over conventional single NVRAM write caches and has great cost advantages over dual-copy NVRAM caches. The RAPID-Cache architecture opens a new dimension for disk system designers to exercise trade-offs among performance, reliability and cost.

Performance of Web servers is critical to the success of many corporations and organizations. However, very few results have been published that quantitatively study the server behavior and identify the performance bottlenecks. In this paper we measured and analyzed the behavior of the popular Apache Web server on a uniprocessor system and a 4-CPU SMP (Symmetric Multi-Processor) system running the IBM AIX operating system. Using the AIX built-in tracing facility, we obtained detailed information on kernel events and system activities while running Apache driven by the SPECweb96 and the WebStone benchmarks. After quantitatively identifying the performance bottlenecks, we proposed and implemented 6 techniques that improve the throughput of Apache by 61%. These techniques are general purpose and can be applied to other Web servers as well.

This paper presents a novel disk storage architecture called DCD, Disk Caching Disk, for the purpose of optimizing I/O performance. The main idea of the DCD is to use a small log disk, referred to as cache-disk, as a secondary disk cache to optimize write performance. While the cache-disk and the normal data disk have the same physical properties, the access speed of the former differs dramatically from the latter because of different data units and different ways in which data are accessed. Our objective is to exploit this speed difference by using the log disk as a cache to build a reliable and smooth disk hierarchy. A small RAM buffer is used to collect small write requests to form a log which is transferred onto the cache-disk whenever the cache-disk is idle. Because of the temporal locality that exists in office/engineering work-load environments, the DCD system shows write performance close to the same size RAM (i.e. solid-state disk) for the cost of a disk. Moreover, the cache-disk can also be implemented as a logical disk in which case a small portion of the normal data disk is used as the log disk. Trace-driven simulation experiments are carried out to evaluate the performance of the proposed disk architecture. Under the office/engineering work-load environment, the DCD shows superb disk performance for writes as compared to existing disk systems. Performance improvements of up to two orders of magnitude are observed in terms of average response time for write operations. Furthermore, DCD is very reliable and works at the device or device driver level. As a result, it can be applied directly to current file systems without the need of changing the operating system.