Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userSystems for Near Real-Time Analysis of Large-Scale Dynamic Graphs
152
0
0.0
(
0
)
Added by
Luis Vaquero
Publication date
2014
fields
Informatics Engineering
and research's language is
English
Ask ChatGPT about the research
No Arabic abstract
Graphs are widespread data structures used to model a wide variety of problems. The sheer amount of data to be processed has prompted the creation of a myriad of systems that help us cope with massive scale graphs. The pressure to deliver fast responses to queries on the graph is higher than ever before, as it is demanded by many applications (e.g. online recommendations, auctions, terrorism protection, etc.). In addition, graphs change continuously (so do the real world entities that typically represent). Systems must be ready for both: near real-time and dynamic massive graphs. We survey systems taking their scalability, real-time potential and capability to support dynamic changes to the graph as driving guidelines. The main techniques and limitations are distilled and categorised. The algorithms run on top of graph systems are not ready for prime time dynamism either. Therefore,a short overview on dynamic graph algorithms has also been included.
rate research
Read More
In cloud storage systems with a large number of servers, files are typically not stored in single servers. Instead, they are split, replicated (to ensure reliability in case of server malfunction) and stored in different servers. We analyze the mean latency of such a split-and-replicate cloud storage system under general sub-exponential service time. We present a novel scheduling scheme that utilizes the load-balancing policy of the textit{power of $d$ $(geq 2)$} choices. An alternative to split-and-replicate is to use erasure-codes, and recently, it has been observed that they can reduce latency in data access (see cite{longbo_delay} for details). We argue that under high redundancy (integer redundancy factor strictly greater than or equal to 2) regime, the mean latency of a coded system is upper bounded by that of a split-and-replicate system (with same replication factor) and the gap between these two is small. We validate this claim numerically under different service distributions such as exponential, shift plus exponential and the heavy-tailed Weibull distribution and compare the mean latency to that of an unsplit-replicated system. We observe that the coded system outperforms the unsplit-replication system by at least $20%$. Furthermore, we consider the mean latency for an erasure coded system with low redundancy (fractional redundancy factor between 1 and 2), a scenario which is more pragmatic, given the storage constraints (cite{rashmi_thesis}). However under this regime, we restrict ourselves to the special case of exponential service time distribution and use the randomized load balancing policy namely textit{batch-sampling}. We obtain an upper bound on mean delay that depends on the order statistics of the queue lengths, which, we further smooth out via a discrete to continuous approximation.
This paper describes a comprehensive prototype of large-scale fault adaptive embedded software developed for the proposed Fermilab BTeV high energy physics experiment. Lightweight self-optimizing agents embedded within Level 1 of the prototype are responsible for proactive and reactive monitoring and mitigation based on specified layers of competence. The agents are self-protecting, detecting cascading failures using a distributed approach. Adaptive, reconfigurable, and mobile objects for reliablility are designed to be self-configuring to adapt automatically to dynamically changing environments. These objects provide a self-healing layer with the ability to discover, diagnose, and react to discontinuities in real-time processing. A generic modeling environment was developed to facilitate design and implementation of hardware resource specifications, application data flow, and failure mitigation strategies. Level 1 of the planned BTeV trigger system alone will consist of 2500 DSPs, so the number of components and intractable fault scenarios involved make it impossible to design an `expert system that applies traditional centralized mitigative strategies based on rules capturing every possible system state. Instead, a distributed reactive approach is implemented using the tools and methodologies developed by the Real-Time Embedded Systems group.
Exploratory data analysis tools must respond quickly to a users questions, so that the answer to one question (e.g. a visualized histogram or fit) can influence the next. In some SQL-based query systems used in industry, even very large (petabyte) datasets can be summarized on a human timescale (seconds), employing techniques such as columnar data representation, caching, indexing, and code generation/JIT-compilation. This article describes progress toward realizing such a system for High Energy Physics (HEP), focusing on the intermediate problems of optimizing data access and calculations for query sized payloads, such as a single histogram or group of histograms, rather than large reconstruction or data-skimming jobs. These techniques include direct extraction of ROOT TBranches into Numpy arrays and compilation of Python analysis functions (rather than SQL) to be executed very quickly. We will also discuss the problem of caching and actively delivering jobs to worker nodes that have the necessary input data preloaded in cache. All of these pieces of the larger solution are available as standalone GitHub repositories, and could be used in current analyses.
With widespread advances in machine learning, a number of large enterprises are beginning to incorporate machine learning models across a number of products. These models are typically trained on shared, multi-tenant GPU clusters. Similar to existing cluster computing workloads, scheduling frameworks aim to provide features like high efficiency, resource isolation, fair sharing across users, etc. However Deep Neural Network (DNN) based workloads, predominantly trained on GPUs, differ in two significant ways from traditional big data analytics workloads. First, from a cluster utilization perspective, GPUs represent a monolithic resource that cannot be shared at a fine granularity across users. Second, from a workload perspective, deep learning frameworks require gang scheduling reducing the flexibility of scheduling and making the jobs themselves inelastic to failures at runtime. In this paper we present a detailed workload characterization of a two-month long trace from a multi-tenant GPU cluster in a large enterprise. By correlating scheduler logs with logs from individual jobs, we study three distinct issues that affect cluster utilization for DNN training workloads on multi-tenant clusters: (1) the effect of gang scheduling and locality constraints on queuing, (2) the effect of locality on GPU utilization, and (3) failures during training. Based on our experience running a large-scale operation, we provide design guidelines pertaining to next-generation cluster schedulers for DNN training workloads.
Simulations and runtime measurements are some of the methods which can be used to evaluate whether a given NoC-based platform can accommodate application workload and fulfil its timing requirements. Yet, these techniques are often time-consuming, and hence can evaluate only a limited set of scenarios. Therefore, these approaches are not suitable for safety-critical and hard real-time systems, where one of the fundamental requirements is to provide strong guarantees that all timing requirements will always be met, even in the worst-case conditions. For such systems the analytic-based real-time analysis is the only viable approach. In this paper the focus is on the real-time communication analysis for wormhole-switched priority-preemptive NoCs. First, we elaborate on the existing analysis and identify one source of pessimism. Then, we propose an extension to the analysis, which efficiently overcomes this limitation, and allows for a less pessimistic analysis. Finally, through a comprehensive experimental evaluation, we compare the newly proposed approach against the existing one, and also observe how the trends change with different traffic parameters.
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
