# Cost Model for Pregel on Many distributed graph computing (DGC) systems like PowerGraph [4] and...

date post

23-May-2020Category

## Documents

view

0download

0

Embed Size (px)

### Transcript of Cost Model for Pregel on Many distributed graph computing (DGC) systems like PowerGraph [4] and...

Cost Model for Pregel on GraphX

Rohit Kumar1,2, Alberto Abelló2, and Toon Calders1,3

1Department of Computer and Decision Engineering Université Libre de Bruxelles, Belgium

2Department of Service and Information System Engineering Universitat Politécnica de Catalunya (BarcelonaTech), Spain

3Department of Mathematics and Computer Science Universiteit Antwerpen, Belgium

Abstract. The graph partitioning strategy plays a vital role in the over- all execution of an algorithm in a distributed graph processing system. Choosing the best strategy is very challenging, as no one strategy is al- ways the best fit for all kinds of graphs or algorithms. In this paper, we help users choosing a suitable partitioning strategy for algorithms based on the Pregel model by providing a cost model for the Pregel implemen- tation in Spark-GraphX. The cost model shows the relationship between four major parameters: 1) input graph 2) cluster configuration 3) algo- rithm properties and 4) partitioning strategy. We validate the accuracy of the cost model on 17 different combinations of input graph, algorithm, and partition strategy. As such, the cost model can serve as a basis for yet to be developed optimizers for Pregel.

1 Introduction

Large graphs with millions of nodes and billions of edges are becoming quite common now. Social media graphs, road network graphs, and relationship graphs between buyers and products are some of the examples of large graphs gener- ated and processed regularly [3]. With the increase in size of these graphs, the classical approach of graph processing is becoming insufficient [7]. Hence, to address these shortcomings, vertex-centric programming models [9] have been proposed to transform the way graph problems are managed. Pregel [10] is one such programming models which supports distributed (parallel) graph compu- tations. Many distributed graph computing (DGC) systems like PowerGraph [4] and Spark-GraphX [14] provide implementations of the Pregel model for graph computations. DGC systems distribute the graph computation by partitioning the graph over different nodes of a cluster.

There are many partitioning strategies proposed in literature [13, 11, 4] for performing efficient graph computations on DGC systems. Most of the DGC sys- tems provide the same programing model and offer similar features and strategies to use. Depending on the internal implementation of these strategies and algo- rithms, the systems can give different performance. Even once a user has decided

a system to use, there are not enough guidelines on which partitioning strategy to use for which application or graph. Verma et.al. in [12] attempts to address this question with an experimental comparison of different partitioning strate- gies on three different DGC systems resulting in a set of rules. However, there is no clear theoretical justification of why one partitioning strategy performs bet- ter than another depending on a particular combination of graph and algorithm. Moreover, the paper does not consider the cluster properties which according to our cost model, is one of the parameters in deciding the best partitioning strategy. In this paper, we address this question by providing a cost model for the Pregel implementation in GraphX. Cost models are used in the database community for query plan evaluation. We contend that DGC systems should be able to choose the best partitioning strategy for a given graph and algorithm using our cost model in iterative graph computations.

Concretely, in this paper, we make the following contributions: (i) we for- mulate a cost model to capture the different dominating factors involved in the Pregel model (Section 3); (ii) we validate our cost model on GraphX by estimat- ing the computation time and comparing it with real execution time (Section 4). To the best of our knowledge this is the first work in which a cost model based approach has been proposed for Pregel to help users to choose the best parti- tioning strategy. Similar cost models could be obtained for Pregel on other DGC systems.

2 Background

In this section, we present background information on (1) the Pregel model, and (2) the different partitioning strategies we used in the experiments.

2.1 Pregel Model

In order to render graph computations more efficient, new graph programming models such as Pregel have been introduced [10]. In Pregel, graph algorithms are expressed as iterative vertex-centric computations which can be easily and transparently distributed automatically. We illustrate this principle with the fol- lowing graph algorithm CC for computing connected components in a graph: we start with assigning to each vertex a unique identifier. In the first step each ver- tex sends a message with its unique identifier to all its neighbors. Subsequently, for each vertex the minimum is computed of all incoming identifiers. If this min- imum is lower than its own identifier, the vertex updates its internal state with this new minimum and sends a message to its neighbors to notify them of its new minimum. This process continues until no more messages are sent. It is easy to see that this iteration will terminate and that the result will be that each vertex holds the minimal identifier over all vertices in its connected component, which can then serve as an identifier of that connected component.

As we can see in this example, a user of Pregel only has to provide the following components:

Fig. 1. An example of Pregel model consisting of three vertices.

– Initialization: one initial message per vertex. In the case of CC, this initial message contains the unique identifier of that vertex;

– Function to combine all incoming messages for a vertex. In our example, the combine function takes the minimum over all incoming identifiers.

– A function called the vertex program to update the internal state of the vertex if the minimum identifier received is less than the current identifier of the vertex.

– A function to send the vertex current identifier to its neighbors. In CC, the internal state of a vertex is updated only if the vertex receives a identifier smaller than it is already storing. Only in that case messages are sent to its neighbors with this updated minimum.

Figure 1 illustrate this programming model; every iteration of running the vertex program and combining the messages that will be input for the next iteration is called a super-step. In the first super-step every vertex is activated and executes its vertex program. In Figure 1, the vertex programs are called “tasks” and the blue lines represent messages sent between vertices. In the second super-step in this figure, vertex 1 does not receive any message and hence will not be active in super-step 2. Vertex 2 receives two messages which are combined and the vertex program is executed. Similarly, vertex 3 receives one message and executes its vertex program. The time it takes for each task could be different and hence there is a synchronization barrier after every super-step. Finally, in super-step 4 no messages are generated and the computation stops.

The main benefit of the Pregel programming model is that it provides a powerful language in which many graph algorithms can be expressed in a natural way. At the same time, however, the programs are flexible enough to allow for automatically and transparently distributing their execution as we will see in next section.

2.2 Partitioning

There are two kinds of partitioning strategies for distributed graph processing: 1) vertex-cut [4] and 2) edge-cut [6, 1]. In vertex-cut partitioning the edges are assigned to partitions and thus the vertices can span partitions i.e vertices are replicated or mirrored across partitions. In edge-cut, the vertices are partitioned and the edge can span across partitions i.e edge is replicated or mirrored across partitions. GraphX utilizes the vertex-cut partitioning strategy. In vertex-cut partitioning, the goal of a partitioning strategy is to partition the edges such that the load (number of edges) in every partition is balanced and vertex repli- cation (number of mirrors of vertex) is minimum. Average replication factor is a common metric to measure the effectiveness of vertex-cut partitioning.

The simplest vertex-cut partitioning strategy is to partition edges using a hash function. GraphX [14] has two different variants for this: Random Vertex Cut (RVC) and Canonical Random Vertex Cut (CRVC). Given a hash function h, RVC assigns an edge (u, v) based on the hash of the source and destination vertex (i.e. A(u, v) = h(u, v) mod k). CRVC partitions the edge regardless of the direction and hence an edge (u, v) and (v, u) will be assigned to the same partition. CRVC or RVC provides a good load balance due to the randomness in assigning the edges but do not grantee any upper bound on the replication factor. There is another strategy which uses two-dimensional sparse matrix and is sim- ilar to grid partitioning [5], EdgePartition2D [2]. In EdgePartition2D partitions are arranged as a square matrix, and for an edge it picks a partition by choosing column on the basis of the hash of the source vertex and row on the basis of the hash of the destination vertex. It ensures a replication factor of (2

√ N − 1)

where N is the number of partitions. In practice, these approaches result in large number of vertex replications and do not perform well for a power-law graphs.

Recently, a Degree-Based Hashing (DBH) algorithm [13] was introduced with improved grantees on replication factor for power-law graphs. DBH partitions edges based on the hash of its lowest degree end point thus forcing replication of high degree