ترغب بنشر مسار تعليمي؟ اضغط هنا

Download time analysis for distributed storage systems with node failures

71   0   0.0 ( 0 )
 نشر من قبل Tejas Bodas
 تاريخ النشر 2021
  مجال البحث الهندسة المعلوماتية
والبحث باللغة English




اسأل ChatGPT حول البحث

We consider a distributed storage system which stores several hot (popular) and cold (less popular) data files across multiple nodes or servers. Hot files are stored using repetition codes while cold files are stored using erasure codes. The nodes are prone to failure and hence at any given time, we assume that only a fraction of the nodes are available. Using a cavity process based mean field framework, we analyze the download time for users accessing hot or cold data in the presence of failed nodes. Our work also illustrates the impact of the choice of the storage code on the download time performance of users in the system.



قيم البحث

اقرأ أيضاً

The paper presents techniques for analyzing the expected download time in distributed storage systems that employ systematic availability codes. These codes provide access to hot data through the systematic server containing the object and multiple r ecovery groups. When a request for an object is received, it can be replicated (forked) to the systematic server and all recovery groups. We first consider the low-traffic regime and present the close-form expression for the download time. By comparison across systems with availability, maximum distance separable (MDS), and replication codes, we demonstrate that availability codes can reduce download time in some settings but are not always optimal. In the high-traffic regime, the system consists of multiple inter-dependent Fork-Join queues, making exact analysis intractable. Accordingly, we present upper and lower bounds on the download time, and an M/G/1 queue approximation for several cases of interest. Via extensive numerical simulations, we evaluate our bounds and demonstrate that the M/G/1 queue approximation has a high degree of accuracy.
Distributed storage systems provide reliable access to data through redundancy spread over individually unreliable nodes. Application scenarios include data centers, peer-to-peer storage systems, and storage in wireless networks. Storing data using a n erasure code, in fragments spread across nodes, requires less redundancy than simple replication for the same level of reliability. However, since fragments must be periodically replaced as nodes fail, a key question is how to generate encoded fragments in a distributed way while transferring as little data as possible across the network. For an erasure coded system, a common practice to repair from a node failure is for a new node to download subsets of data stored at a number of surviving nodes, reconstruct a lost coded block using the downloaded data, and store it at the new node. We show that this procedure is sub-optimal. We introduce the notion of regenerating codes, which allow a new node to download emph{functions} of the stored data from the surviving nodes. We show that regenerating codes can significantly reduce the repair bandwidth. Further, we show that there is a fundamental tradeoff between storage and repair bandwidth which we theoretically characterize using flow arguments on an appropriately constructed graph. By invoking constructive results in network coding, we introduce regenerating codes that can achieve any point in this optimal tradeoff.
We present Kaleidoscope an innovative system that supports live forensics for application performance problems caused by either individual component failures or resource contention issues in large-scale distributed storage systems. The design of Kale idoscope is driven by our study of I/O failures observed in a peta-scale storage system anonymized as PetaStore. Kaleidoscope is built on three key features: 1) using temporal and spatial differential observability for end-to-end performance monitoring of I/O requests, 2) modeling the health of storage components as a stochastic process using domain-guided functions that accounts for path redundancy and uncertainty in measurements, and, 3) observing differences in reliability and performance metrics between similar types of healthy and unhealthy components to attribute the most likely root causes. We deployed Kaleidoscope on PetaStore and our evaluation shows that Kaleidoscope can run live forensics at 5-minute intervals and pinpoint the root causes of 95.8% of real-world performance issues, with negligible monitoring overhead.
510 - Mohamed Kaaniche 2007
We present a hierarchical simulation approach for the dependability analysis and evaluation of a highly available commercial cache-based RAID storage system. The archi-tecture is complex and includes several layers of overlap-ping error detection and recovery mechanisms. Three ab-straction levels have been developed to model the cache architecture, cache operations, and error detection and recovery mechanism. The impact of faults and errors oc-curring in the cache and in the disks is analyzed at each level of the hierarchy. A simulation submodel is associated with each abstraction level. The models have been devel-oped using DEPEND, a simulation-based environment for system-level dependability analysis, which provides facili-ties to inject faults into a functional behavior model, to simulate error detection and recovery mechanisms, and to evaluate quantitative measures. Several fault models are defined for each submodel to simulate cache component failures, disk failures, transmission errors, and data errors in the cache memory and in the disks. Some of the parame-ters characterizing fault injection in a given submodel cor-respond to probabilities evaluated from the simulation of the lower-level submodel. Based on the proposed method-ology, we evaluate and analyze 1) the system behavior un-der a real workload and high error rate (focusing on error bursts), 2) the coverage of the error detection mechanisms implemented in the system and the error latency distribu-tions, and 3) the accumulation of errors in the cache and in the disks.
Memory-bound algorithms show complex performance and energy consumption behavior on multicore processors. We choose the lattice-Boltzmann method (LBM) on an Intel Sandy Bridge cluster as a prototype scenario to investigate if and how single-chip perf ormance and power characteristics can be generalized to the highly parallel case. First we perform an analysis of a sparse-lattice LBM implementation for complex geometries. Using a single-core performance model, we predict the intra-chip saturation characteristics and the optimal operating point in terms of energy to solution as a function of implementation details, clock frequency, vectorization, and number of active cores per chip. We show that high single-core performance and a correct choice of the number of active cores per chip are the essential optimizations for lowest energy to solution at minimal performance degradation. Then we extrapolate to the MPI-parallel level and quantify the energy-saving potential of various optimizations and execution modes, where we find these guidelines to be even more important, especially when communication overhead is non-negligible. In our setup we could achieve energy savings of 35% in this case, compared to a naive approach. We also demonstrate that a simple non-reflective reduction of the clock speed leaves most of the energy saving potential unused.
التعليقات
جاري جلب التعليقات جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
mircosoft-partner

هل ترغب بارسال اشعارات عن اخر التحديثات في شمرا-اكاديميا