Packet based Network-on-Chip (NoC) connect tens to hundreds of components in a multi-core system. The routing and arbitration policies employed in traditional NoCs treat all application packets equally. However, some packets are critical as they stall application execution whereas others are not. We differentiate packets based on a metric called slack that captures a packet's criticality. We observe that majority of NoC packets generated by standard application based benchmarks do not have slack and hence are critical. Prioritising these critical packets during routing and arbitration will reduce application stall and improve performance. We study the diversity and interference of packets to propose a policy that prioritises critical packets in NoC. This paper presents a slack-aware re-routing (SAR) technique that prioritises lower slack packets over higher slack packets and explores alternate minimal path when two no-slack packets compete for same output port. Experimental evaluation on a 64-core Tiled Chip Multi-Processor (TCMP) with 8×8 2D mesh NoC using both multiprogrammed and multithreaded workloads show that our proposed policy reduces application stall time by upto 22% over traditional round-robin policy and 18% over state-of-the-art slack-aware policy.

Critical packet prioritisation by slack-aware re-routing in on-chip networks

The elasticity feature of cloud computing has been proved as pertinent for parallel applications, since users do not need to take care about the best choice for the number of processes/resources beforehand. To accomplish this, the most common approaches use threshold-based reactive elasticity or time-consuming proactive elasticity. However, both present at least one problem related to: the need of a previous user experience, lack on handling load peaks, completion of parameters or design for a specific infrastructure and workload setting. In this regard, we developed a hybrid elasticity service for Master-Slave parallel applications named Helpar (Hybrid Elasticity Model for Parallel Applications). As parameterless model, Helpar presents a closed control loop elasticity architecture that adapts at runtime the values of lower and upper thresholds. Thus, we intend to provide a practical and effortless realization of the cloud elasticity and parallel computing duet, so delivering this capability as a plug-and-play utility to end users. Besides presenting Helpar, our purpose is to provide a comparison between Helpar and our previous work on reactive elasticity called AutoElastic. We will explore different metrics, including applications’ time, energy consumption and cost, as well as distinct types of workloads when executing a scientific HPC application. The results present the Helpar’s lightweight feature, besides highlighting its performance competitiveness in terms of application time and cost (performance × energy) metrics. In other words, the hand-tuning of thresholds in AutoElastic often is responsible for the best results, but this procedure may be time-consuming besides optimized for a particular set of application and infrastructure.

Towards Enabling Live Thresholding as Utility to Manage Elastic Master-Slave Applications in the Cloud

With the rapid development of networks and Information technologies, cloud computing is not only becoming popular, the types of cloud services available are also increasing. Through cloud services, users can upload their requirements via the Internet to the cloud environment and receive responses following post-processing, for example, with cloud storage services. Software-Defined Storage (SDS) is a virtualization technology for cloud storage services. SDS uses software to integrate storage resources and to improve the accessibility and usability of storage services. Currently, there are many different open source projects available for SDS development. This work aims to utilize these open source projects to improve the efficiency of integration for hardware and software resources. In other words, in this work, we propose a cloud storage system that integrates various open source SDS software to make cloud storage services more compatible and user friendly. The cloud service systems can also be managed in a more convenient and flexible manner. The experimental results demonstrate the benefits of the proposed system.

/pdf/on-construction-of-a-cloud-storage-system-with-heterogeneous-215na1xcnm.pdf

On construction of a cloud storage system with heterogeneous software-defined storage technologies

In enterprise data centers, reliable performance models on storage devices are desirable for efficient storage management and optimization. However, many cloud environments consist of heterogeneous storage devices, e.g., a mixture of commodity disks, where accurate performance models are of particular challenge to attain. In this paper, we propose a lightweight queueing-based storage performance modeling framework, which is able to infer the maximum IO load that a storage device can sustain, as well as its IO load v.s. response time performance curve. Our inference framework views the underlying storage resources as blackboxes and only utilizes historical measurements of the IO and response time on the devices. In an OpenStack environment, we also develop a new storage volume placement algorithm using our performance inference and modeling framework. Experimental results show that our solution can provide up to 80% increase of the IO throughput, in tandem with a 40% reduction of the average response time, compared to the performance provided by the default OpenStack policy.

Queueing-based storage performance modeling and placement in OpenStack environments

This paper presents a novel metadata management mechanism on the metadata server (MDS) for parallel and distributed file systems. In this technique, the client file system backs up the sent metadata requests, which have been handled by the metadata server, so that the MDS does not need to log metadata changes to nonvolatile storage for achieving highly available metadata service, as well as better performance improvement in metadata processing. As the client file system backs up certain sent metadata requests in its memory, the overhead for handling these backup requests is much smaller than that brought by the metadata server, while it adopts logging or journaling to yield highly available metadata service. The experimental results show that this newly proposed mechanism can significantly improve the speed of metadata processing and render a better I/O data throughput, in contrast to conventional metadata management schemes, that is, logging or journaling on MDS. Besides, a complete metadata recovery can be achieved by replaying the backup logs cached by all involved clients, when the metadata server has crashed or gone into nonoperational state exceptionally.

/pdf/log-less-metadata-management-on-metadata-server-for-parallel-10cfwx06y5.pdf

Log-Less Metadata Management on Metadata Server for Parallel File Systems

We explore the un-fairness problem when heterogeneous applications share the cache.A dynamic partitioning caching algorithm is proposed for fairness and performance.We use both theory analysis and experiments to support our algorithm.The experiments show that our algorithm gets significant improvement in fairness.The experiments show that our algorithm get slightly improvement in performance. In this paper, we investigate the problem of fair storage cache allocation among multiple competing applications with diversified access rates. Commonly used cache replacement policies like LRU and most LRU variants are inherently unfair in cache allocation for heterogeneous applications. They implicitly give more cache to the applications that has high access rate and less cache to the applications of slow access rate. However, applications of fast access rate do not always gain higher performance from the additional cache blocks. In contrast, the slow application suffer poor performance with a reduced cache size. It is beneficial in terms of both performance and fairness to allocate cache blocks by their utility.In this paper, we propose a partition-based cache management algorithm for a shared cache. The goal of our algorithm is to find an allocation such that all heterogeneous applications can achieve a specified fairness degree as least performance degradation as possible. To achieve this goal, we present an adaptive partition framework, which partitions the shared cache among competing applications and dynamically adjusts the partition size based on predicted utility on both fairness and performance. We implement our algorithm in a storage simulator and evaluate the fairness and performance with various workloads. Experimental results show that, compared with LRU, our algorithm achieves large improvement in fairness and slightly in performance.

Heterogeneous-aware cache partitioning

Modern storage environment is commonly composed of heterogeneous storage devices. However, traditional cache algorithms exhibit performance degradation in heterogeneous storage systems because they were not designed to work with the diverse performance characteristics. In this paper, we present a new cache algorithm called HCM for heterogeneous storage systems. The HCM algorithm partitions the cache among the disks and adopts an effective scheme to balance the work across the disks. Furthermore, it applies benefit-cost analysis to choose the best allocation of cache block to improve the performance. Conducting simulations with a variety of traces and a wide range of cache size, our experiments show that HCM significantly outperforms the existing state-of-the-art storage-aware cache algorithms.

/pdf/an-effective-cache-algorithm-for-heterogeneous-storage-24rapfkv3u.pdf

An Effective Cache Algorithm for Heterogeneous Storage Systems

In this paper, we investigate the problem of fair storage cache allocation among multiply competing applications with diversified access rates. Commonly used cache replacement policies like LRU and most LRU variants are inherently unfair in cache allocation for heterogenous applications. They implicitly give more cache to the applications that has high access rate and less cache to the applications of slow access rate. However, applications of fast access rate do not always gain higher performance from the additional cache blocks. In contrast, the slow application suffer poor performance with a reduced cache size. It is beneficial in terms of both performance and fairness to allocate cache blocks by their utility.In this paper, we propose a partition-based cache management algorithm for a shared cache. The goal of our algorithm is to find an allocation such that all heterogenous applications can achieve a specified fairness degree, while maximizing the overall performance. To achieve this goal, we present an adaptive partition framework, which partitions the shared cache among competing applications and dynamic adjusts the partition size based on predicted utility on both fairness and performance. We implemented our algorithm in a storage simulator and evaluated the fairness and performance with various workloads. Experimental results show that, compared with LRU, our algorithm achieves large improvement in fairness and slightly in performance.

Enhancing both fairness and performance using rate-aware dynamic storage cache partitioning

This work presents a new black-box learning-based model to predict the performance of disk array We focus on predicting how the change of performance with comprised different disks in a disk array We use CART-MARS hybrid models: the CART methods to model the correlation between the workload and disk array The MARS methods to construct the delta model for pairs of disk arrays Compared with direct prediction of performance, the delta model avoids the complexity of modeling the internal architecture and algorithms of a disk array system Our experiments show that the error ratio of the prediction of delta model is reduced to 16% By comparison, the error ratio of the direct prediction is 29%, which is almost two times larger than delta model

Disk array performance prediction with CART-MARS hybrid models

Modern storage systems are becoming increasingly consolidated. As a result, there exists a competition for resources among concurrent streams. Sequential prefetching is widely used in modern storage system. Most of them always ignore the impact of diversified access rate of concurrent streams. However the concurrent streams with diversified access rate will introduce several problems. The streams with fast access rate may evict the cache blocks of the streams with slow access rate and lead the slow streams re-fetch the evicted cache blocks in future access, which makes the prefetching wastage and unfairness to the slow streams. We design and implement a novel, adaptive algorithm, named ASPM to solve these problems. Our experiments show that, compared with LRU and AMP, ASPM can achieve significantly improvement in fairness among concurrent streams and slightly in performance (average on 6.8%).

Yong Li

Papers

Heterogeneous-aware cache partitioning

An Effective Cache Algorithm for Heterogeneous Storage Systems

Enhancing both fairness and performance using rate-aware dynamic storage cache partitioning

Disk array performance prediction with CART-MARS hybrid models

Adaptive Sequential Prefetching for Multiple Streams