Citation Context

...g interface that can provide fault tolerance efficiently. Existing abstractions for in-memory storage on clusters, such as distributed shared memory [24], keyvalue stores [25], databases, and Piccolo =-=[27]-=-, offer an interface based on fine-grained updates to mutable state (e.g., cells in a table). With this interface, the only ways to provide fault tolerance are to replicate the data across machines or...

Citation Context

...ge. As of the writing of this paper, public implementations of Pregel (e.g., Giraph) were unable to handle even our smaller synthetic problems due to memory limitations. Consequently, we used Piccolo =-=[32]-=- as a proxy implementation of Pregel since Piccolo naturally expresses the Pregel abstraction and provides an efficient C++ implementation with dynamic load-balancing. Finally, we used our implementat...

Citation Context

... computation model, in which the user defines a program that is executed locally for each vertex in parallel. In addition, high-performance systems that are based on key-value tables, such as Piccolo =-=[40]-=- and Spark [48], can efficiently represent many graph-parallel algorithms. Current graph systems are able to scale to graphs of billions of edges by distributing the computation. However, while distri...

Citation Context

...puting: fault tolerance, scheduling, synchronisation and communication. MapReduce and Dryad can be used to implement a wide range of algorithms [3, 39], but they are awkward or inefficient for others =-=[12, 21, 25, 28, 34]-=-. The problems typically arise with iterative algorithms, which underlie many machine-learning and optimisation problems, but require a more expressive programming model and a more powerful execution ...

Citation Context

...use of relatively simple eviction policies and do not coordinate scheduling with locality since they were designed for usage by a network of workstations. Cluster computing frameworks such as Piccolo =-=[25]-=- and Spark [21] are optimized for iterative machine learning workloads. They cache data in memory after the first iteration, speeding up further iterations. The key difference with PACMan is that sinc...

Citation Context

...since the StateTable supports random access, it is not necessary to perform sort between the map and reduce in PrIter, so that we eliminate the sort phase, which can significantly improve performance =-=[17]-=-. Moreover, we start the reduce operation immediately upon receiving a map’s output. In other words, the “reduce” function is applied on 〈key, value〉 rather than 〈key, values list〉. It updates the cor...

Citation Context

...the computation. To better support various graph mining algorithms that might require different inter-vertex communication patterns, Kineograph support the push and the pull models in the computation =-=[17, 26]-=-. To check the status of a vertex and perform computation, the push model allows other vertices to push updates to a specific vertex and a vertex in the pull model proactively pulls data from neighbor...

Citation Context

...nication), (2) supports inference for a broad collection of ML methods, and (3) enjoys correctness guarantees. In this paper, we explore a path to such a system using the 1 idea of a parameter server =-=[22, 2]-=-, which we define as the combination of a shared key-value store that provides a centralized storage model (which may be implemented in a distributed fashion) with a synchronization model for reading/...

Citation Context

...However, the two orders of magnitude more disk than memory capacity in many a typical data centers implies that we use memory as only a cache. Iterative Jobs New cluster computing frameworks (Piccolo =-=[29]-=-, and Spark [27]) are optimized to support machine learning applications and interactive data analysis, that iteratively read the same data. While the first iteration reads data from disk, these frame...