Novel Parallel Algorithms for Fast MultiGPUBased Generation of Massive ScaleFree Networks
 169 Downloads
Abstract
A novel parallel algorithm is presented for generating random scalefree networks using the preferential attachment model. The algorithm, named cuPPA, is customdesigned for “single instruction multiple data (SIMD)” style of parallel processing supported by modern processors such as graphical processing units (GPUs). To the best of our knowledge, our algorithm is the first to exploit GPUs, and also the fastest implementation available today, to generate scalefree networks using the preferential attachment model. A detailed performance study is presented to understand the scalability and runtime characteristics of the cuPPA algorithm. Also another version of the algorithm called cuPPAHash tailored for multiple GPUs is presented. On a single GPU, the original cuPPA algorithm delivers the best performance, but is challenging to port to multiGPU implementation. For multiGPU implementation, cuPPAHash has been used as the parallel algorithm to achieve a perfect linear speedup up to 4 GPUs. In one of the best cases, when executed on an NVidia GeForce 1080 GPU, the original cuPPA generates a scalefree network of two billion edges in less than 3 s. On multiGPU platforms, cuPPAHash generates a scalefree network of 16 billion edges in less than 7 s using a machine consisting of 4 NVidia Tesla P100 GPUs.
Keywords
GPU Preferential attachment Random networks Scalefree networks1 Introduction
Networks are prevalent in many complex systems such as circuits, chemical compounds, protein structures, biological networks, social networks, the Web, and XML documents. Recently, there has been substantial interest in the study of a variety of random networks to serve as mathematical models of complex systems. Various network theories, metrics, topology, and mathematical models have been proposed to understand the underlying properties and relationships of these systems. Among the proposed network models, the first and the most studied model is the Erdős–Rényi model [14]. However, the Erdős–Rényi model does not exhibit the characteristics observed in many realworld complex systems [8]. Barabási and Albert [8] discovered a class of inhomogeneous networks, called scalefree networks, characterized by a powerlaw degree distribution \(P(k)\propto k^{\gamma }\), where k represents the degree of a vertex and \(\gamma\) is a constant. While high degree vertices are improbable in Erdős–Rényi networks, they do occur with statistically significant probability in scalefree networks. Furthermore, the work of Albert et al. [6] suggests these high degree vertices appear to play an important role in the behavior of scalefree systems, particularly with respect to their resilience [11]. For example, the Barabasi–Albert model can be used for evaluating the North American electric grid with high reliability [11].
As these complex systems of today grow larger, the ability to generate progressively large random networks becomes all the more important. It is well known that the structure of larger networks is fundamentally different from that of small networks, and many patterns such as communities emerge only in massive networks [20]. Although various random network models have been used and studied over the last several decades, even efficient sequential algorithms for generating such networks were nonexistent until recently. The efficient sequential algorithms are able to generate networks with millions of edges in a reasonable amount of time; however, generating networks with billions of edges can take a prohibitively large amount of time. This motivates the need for efficient parallel algorithms for generating such networks. Naïve parallelization of the sequential algorithms for generating random networks may not work due to the dependencies among the edges and the possibility of creating duplicate (parallel) edges.
One of the earliest known parallel algorithms for the preferential attachment model is given by Yoo and Henderson [26]. Although useful, the algorithm has two weaknesses: (1) for ease of handling dependencies and avoid the required complex synchronization, they adopted an approximation algorithm rather than an exact algorithm; and (2) to correctly generate the network, the algorithm needs manual adjustment of several control parameters. An exact distributedmemory parallel algorithm was presented in [4]. A literature review of the recent developments is presented in Sect. 6.
Graphics processors (GPUs) are a costeffective, energyefficient, and widely available parallel processing platform. GPUs are highly parallel, multithreaded, manycore processors that have greatly expanded beyond graphics operations and are now widely used for general purpose computing. The use of GPUs is prevalent in many areas such as scientific computation, complex simulations, big data analytics, machine learning, and data mining. However, there is a lack of GPUbased graph/network generators, especially for scalefree networks such as those based on the preferential attachment model. There exist GPUbased network generators for Erdős–Rényi networks [23] and smallworld model [18]. However, until recently we found no GPUbased algorithm to generate scalefree networks [3]. In this paper, we present cuPPA, a novel GPUbased algorithm for generating networks conforming to the preferential attachment model. The algorithm adopts the copy model [17] and employs a simpler synchronization technique suitable for GPUs. With cuPPA, one can generate a network with two billion edges using a modern NVidia GPU in less than 3 s. To the best of our knowledge, this is the first GPUbased algorithm to generate networks using the exact preferential attachment model. Although cuPPA works really well on a single GPU, generating bigger networks with multiple GPUs is a challenging issue due to the complex synchronization and message communication required among the GPUs. We present another algorithm called cuPPAHash to generate networks using preferential attachment model on multiple GPUs. The algorithm uses hashing instead of pseudorandom number generators and does not require any communication among the GPUs. With cuPPAHash, we generated a network of 16 billion edges in less than 7 s using a machine consisting of 4 NVidia Tesla P100 GPUs.
The rest of the paper is organized as follows: In Sect. 2, background material is provided in terms of preliminary information, notations, an outline of the network generation problem, and two leading sequential algorithms. In Sect. 3, our parallel cuPPA algorithm for the GPU is presented. In Sect. 4, we present a multiGPU algorithm called cuPPAHash. The experimental study and performance results using cuPPA are described in Sect. 5. We present a review of related works in Sect. 6. Finally, Sect. 7 concludes with a summary and an outline of future directions.
2 Background
2.1 Preliminaries and Notations
In the rest of this paper, we use the following notations. We denote a network G(V, E), where V and E are the sets of vertices and edges, respectively, with \(m = E\) edges and \(n = V\) vertices labeled as \(0, 1, 2, \dots , n1\). For any \((u,v)\in E\), we say u and v are neighbors of each other. The set of all neighbors of \(v \in V\) is denoted by N(v), i.e., \(N(v)=\{u \in V  (u,v) \in E\}\). The degree of v is \(d_v = N(v)\). If u and v are neighbors, sometimes we say that u is connected to v and vice versa.
We develop parallel algorithms using the CUDA (Compute Unified Device Architecture) framework on the GPU. A GPU contains multiple streaming multiprocessors (SMs). An SM is a group of core processors. Each core processor executes only one thread at a time. All core processors can execute their corresponding threads simultaneously. If some threads perform operations that have to wait for data fetches with high latencies, those are put into the waiting state and other pending threads are executed. Therefore, GPUs increase throughput by keeping the processors busy. All thread management, including the creation and scheduling of threads, is performed entirely in hardware with virtually zero overhead and requires negligible time for launching work on the GPU. For these advantages, modern supercomputers such as Summit and Titan, two of the largest supercomputers in the USA, are built using GPUs in addition to conventional central processing units (CPUs).
We use K, M, and B to denote thousand, million, and billion, respectively, e.g., 2 B stands for two billion.
2.2 Preferential AttachmentBased Models
The preferential attachment model is a model for generating randomly evolved scalefree networks using a preferential attachment mechanism. In a preferential attachment mechanism, a new vertex is added to the network and connected to some existing vertices that are chosen preferentially based on some properties of the vertices. In the most common method, preference is given to vertices with larger degrees: The higher the degree of a vertex, the higher is the probability of choosing it. In this paper, we study only the degreebased preferential attachment, and in the rest of the paper, by preferential attachment (PA) we mean degreebased preferential attachment.
Before presenting our parallel algorithms for generating PA networks, we briefly discuss the sequential algorithms for the same. Many preferential attachmentbased models have been proposed in the literature. Two of the most prominent models are the Barabási–Albert model [8] and the copy model [17] as discussed below.
2.3 Sequential Algorithm: Barabási–Albert Model
One way to generate a random PA network is to use the Barabási–Albert (BA) model. Many realworld networks have two important characteristics: (1) They are evolving in nature and (2) the network tends to be scalefree [8]. In the BA model, a new vertex is connected to an existing vertex that is chosen with probability directly proportional to the current degree of the existing vertex.
The BA model works as follows: Starting with a small clique of \(\hat{d}\) vertices, in every time step, a new vertex t is added to the network and connected to \(d \le \hat{d}\) randomly chosen existing vertices: \(F_\ell (t)\) for \(1 \le \ell \le d\) with \(F_\ell (t) < t\); that is, \(F_\ell (t)\) denotes the \(\ell\)th vertex which t is connected. Thus, each phase adds d new edges \((t, F_1(t)), (t, F_2(t)), \ldots , (t, F_d(t))\) to the network, which exhibits the evolving nature of the model. Let \(\mathbb {F}(t)=\left\{ F_1(t), F_2(t), \ldots , F_d(t) \right\}\) be the set of outgoing vertices from t. Each of the d end points in the set \(\mathbb {F}(t)\) is randomly selected based on the degrees of the vertices in the current network. In particular, the probability \(P_i(t)\) that an outgoing edge from vertex t is connected to vertex \(i < t\) is given by \(P_i(t)=\frac{d_{i}}{\sum _{j}d_{j}}\), where \(d_j\) represents the degree of vertex j.
The networks generated by the BA model are called the BA networks, which bear the aforementioned two characteristics of a realworld network. BA networks have powerlaw degree distribution. A degree distribution is called power law if the probability that a vertex has degree d is given by \(\Pr \left[ d\right] \propto d^{\gamma }\), where \(\gamma \ge 1\) is a positive constant. Barabási and Albert showed that the preferential attachment method of selecting vertices results in a powerlaw degree distribution [8].
A naïve implementation of network generation based on the BA model takes \(\Omega (n^2)\) time where n is the number of vertices. Batagelj and Brandes give an efficient algorithm with a running time of \(\mathcal {O}(m)\) where m is the number of edges [9]. This algorithm maintains a list of vertices such that each vertex i appears in this list exactly \(d_i\) times. The list can easily be updated dynamically by simply appending u and v to the list whenever a new edge (u, v) is added to the network. Now, to find F(t), a vertex is chosen from the list uniformly at random. Since each vertex i occurs exactly \(d_i\) times in the list, we have the probability \(\Pr \left[ F(t) = i\right] = \frac{d_{i}}{\sum _{j }d_{j}}\).
2.4 Sequential Algorithm: Copy Model
As it turns out, the BA model does not easily lend itself to an efficient parallelization [4]. Another algorithm called the copy model [16, 17] preserves preferential attachment and powerlaw degree distribution. The copy model works as follows: Similar to the BA model, it starts with a small clique of \(\hat{d}\) vertices and in every time step, a new vertex t is added to the network to create \(d \le \hat{d}\) connections to existing vertices \(F_\ell (t)\) for \(1 \le \ell \le d\) with \(F_\ell (t) < t\). For each connection \((t,F_\ell (t))\) from vertex t, the following steps are executed:
Step 1: First, a random vertex \(k \in [0, t1]\) is chosen with uniform probability.
We also denote \(\mathbb {F}(t)=\left\{ F_1(t), F_2(t), \ldots , F_d(t) \right\}\) to be the set of outgoing vertices from vertex t.
It can be easily shown that a connection from vertex t to vertex i is made with probability \(\Pr \left[ i \in \mathbb {F}(t) \right] = \frac{d_{i}}{\sum _{j }d_{j}}\) when \(p=\frac{1}{2}\). Thus, when \(p=\frac{1}{2}\), this algorithm follows the Barabási–Albert model as shown in [2, 4].
Thus, the copy model is more general than the BA model. It has been previously shown [17] that the copy model produces networks with degree distribution that follows a powerlaw \(d^{\gamma }\), where the value of the exponent \(\gamma\) depends on the choice of p. Further, it is easy to see the running time of the copy model is \(\mathcal {O}(m)\). The copy model has been used to develop efficient parallel algorithms for generating preferential attachment networks in distributedmemory and sharedmemory machines [4, 7]. In our work presented in this paper, we adopt the copy model as a starting point to design and develop our GPUbased parallel algorithm.
3 GPUbased Parallel Algorithm: cuPPA
The PA model imposes a critical dependency that every new vertex needs to have the state of the previous network to compute its edges. This poses a major challenge in parallelizing preferential attachment algorithms. In phase v, to determine F(v), it requires that \(F_i\) is known for each \(i < v\). As a result, any algorithm for preferential attachment apparently seems to be highly sequential in nature: Phase v cannot be executed until all previous phases are completed.
In [4], a distributedmemorybased algorithm was proposed that exploits the copy model to relieve this sequentiality and run in parallel. We reexamined that exploitation and designed cuPPA, an efficient parallel algorithm for generating preferential attachmentbased networks on a single GPU as described next. Here, we assume that the entire network can be stored in the GPU memory.
Symbols used in this paper
Symbol  Description 

n  The number of vertices 
V  The set of vertices 
m  The number of edges 
E  The set of edges 
T  The number of threads 
d  The number of outgoing edges generated from each new vertex 
p  The probability of creating a direct edge in the copy model 
N(v)  The set of neighbors of vertex v 
\(d_v\)  The degree of vertex v 
\(F_t(k)\)  The outgoing end of kth edge from vertex t 
\(\mathbb {F}_t\)  The set of outgoing ends of edges from vertex t 
\(\mathcal {Q}\)  A queue for the current set of unfinished edges 
\(\mathcal {Q}'\)  A queue for the next set of unfinished edges 
In the first phase (lines 3–21), the algorithm executes the copy model for all of its vertices. The edges that could not be completed are stored in a queue \(\mathcal {Q}\) to be processed later. We call the queue a waiting queue. Each of the other vertices from d to \(n1\) generates d new edges. There are fundamentally two important issues that need to be handled: (1) how we select \(F_{\ell }(v)\) for vertex v where \(1 \le \ell \le d\), and (2) how we avoid duplicate edge creation. Multiple edges for a vertex v are created by repeating the same procedure d times (line 4), and duplicate edges are avoided by simply checking if such an edge already exists—such a check is done whenever a new edge is created.
For the \(\ell\)th edge of a vertex v, another vertex u is chosen from \([1, v1]\) uniformly at random (line 5, 6). Edge (v, u) is created with probability p (line 7). However, before creating such an edge (v, u) in line 8, the existence of such an edge is checked immediately before creating them in line 9. If the edge already exists at that time, the edge is discarded and the process is repeated again (line 5). With the remaining \(1p\) probability, v is connected to some vertex in \(\mathbb {F}(u)\); that is, we make an edge \((v, F_{\ell }(u))\), such that \(\ell\) is chosen from [1, d] uniformly at random.
After the first phase is completed, the algorithm starts to resolve all incomplete edges by processing the waiting queue (lines 22–29). If an item in the current queue \(\mathcal {Q}\) could not be resolved during this step, it is subsequently placed in another queue \(\mathcal {Q}'\). After all incomplete edges on the queue \(\mathcal {Q}\) are processed, the queues \(\mathcal {Q}\) and \(\mathcal {Q}'\) are swapped and \(\mathcal {Q}'\) is cleared. We repeat this process until both the queues are empty. Note that there is no circular dependency in the copy edges. For two vertices \(u > v\), copy edges from u may depend on edges from v but not vice versa. Therefore, there would be no circular waiting and no deadlock situation for the waiting queue to complete.
3.1 Graph Representation
We use one array G of nd elements to represent and store the entire graph. Each vertex u connects to d existing vertices. The neighbors of u are stored between the indices inclusive from ud to \((u+1)d 1\) that represents the other endpoint vertices. We call these indices the outgoing vertex list for vertex u. The initial network consists of the \(d^2\) vertices from the start of the array. For any edge u, v where \(u>v\) and \(u,v>d\), the edge is represented by storing v in one of the d items in the outgoing vertex list of u. Note that the graph G contains exactly nd edges as defined by the Barabási–Albert or the copy model. Any vertex with the index \(0 \le i < nd\) of the array G denotes the \((t \mod d)\)th end point of the vertex \(\frac{i}{d}\).
3.2 Partitioning and Load Balancing
3.3 Segmented RoundRobin Partitioning
In order to exploit the faster shared memory without overflowing the waiting queue capacity, we use a modified roundrobin partitioning scheme called, segmented roundrobin partitioning (SRRP). In this scheme, the entire set of vertices V is first partitioned into some k consecutive subsets \(S_1,S_2,S_3 \ldots S_k\) called segments. From the definition of the copy model, it is clear that vertices on a segment \(S_i\) may only depend on vertices on segment \(S_j\) where \(i \ge j\) but not vice versa. Therefore, the segments have to be processed in a consecutive fashion. Let \(B_i=S_i\) denote the number of elements (also called the segment size) in segment \(S_i\) where \(1 \le i \le k\). Next, the parallel algorithm is executed in k consecutive rounds where round i executes the parallel algorithm for all the vertices in segment \(S_i\). In round i, the \(B_i\) vertices in segment \(S_i\) are further partitioned into T subsets \(V_0(S_i), V_1(S_i), \ldots V_{T1}(S_i),\) using the roundrobin scheme discussed above and executed in parallel using the T threads. The technique is illustrated in Fig. 2.

Fixed Segment Size: The simplest way is to use a fixed sized segments in each round. From the previous discussion, it is clear that in the worst case we need d items per vertex to be placed on the waiting queue. Therefore, we can use up to \(\tau = \min \left( \frac{\mathcal {C}}{d}, \theta \right)\) threads per block where \(\mathcal {C}\) is the total queue capacity and \(\theta\) is the maximum number of threads per block. Then, the segment size is \(\frac{\mathcal {C}}{d\tau }\) vertices per segment. Note that we can exploit the shared memory for \(d \le \mathcal {C}\); otherwise, we need to use the global memory. However, in almost all practical scenarios we have \(d \ll \mathcal {C}\); hence, we can take advantages of the shared memory.

Dynamic Segment Size: Although the fixed segment size scheme ensures that the queue will not overflow in any round, it may not be the most efficient implementation. We use another scheme where the segment size is determined dynamically between two rounds based on the current state of the algorithm. In this scheme, we start with the number of threads per block \(\tau\) and the segment size \(\frac{\mathcal {C}}{d\tau }\) vertices per segment as was done in the fixed segment size scheme. However, at the end of each round, we determine the maximum number of items that were placed in the waiting queue per thread. If the number of items placed in the waiting queue in the round is less than some f factor of the waiting queue capacity per thread \(\frac{\mathcal {C}}{\tau }\), we increase the total capacity \(\mathcal {C}\) by a factor of f (typically, we set \(f=2\)). Before the next round, we recompute the required number of threads per block and update the segment size accordingly.
3.4 CUDASpecific Deadlock Scenario
In the roundrobin scheme, completion of a copy edge of a vertex in a thread \({\mathcal {T}}_i\) may depend on some other thread \({\mathcal {T}}_j\) where \(i \ne j\). Due to the nature of dependency, \({\mathcal {T}}_j\) also may have a copy edge that depends on another vertex that belongs to \({\mathcal {T}}_i\). Therefore, if any of these threads are not running simultaneously on the GPU, the other thread will not be able to complete and a deadlock situation may arise. To avoid such a situation, we must ensure that either all the GPU threads are running concurrently or the dependent threads are put to sleep for a while. In the current CUDA framework, the runtime engine schedules each kernel block to a streaming multiprocessor, and the blocks of running threads are nonpreemptible. Therefore, to ensure that threads are running concurrently to avoid deadlock situation, we cannot use more blocks than the number of available streaming multiprocessors. Note that the upcoming CUDA runtime supports cooperative groups. On such future systems, the deadlock situation could be avoided using block sizes larger than the number of shared multiprocessors.^{1}
4 Generating Networks Using Multiple GPUs
So far we have presented an algorithm to generate the preferential attachmentbased model using a single GPU. Our algorithm works well if the entire network can be stored in the GPU memory. However, the size of the generated network is limited by the amount of GPU memory. Nowadays, it is very common to have multiple GPUs in a computing cluster, even on commodity machines. Therefore, we could potentially use multiple GPUs to generate even larger networks. In this section, we discuss how cuPPA can be extended for multiple GPUs.
Similar to distributing the works of generating edges into multiple threads as discussed at the beginning of the section, we need to distribute the vertices into multiple GPUs. Let the vertices V be partitioned into g subsets \(\mathbb {V}_1,\mathbb {V}_2,\ldots ,\mathbb {V}_g\) where g is the number of available GPUs. GPU i processes the edges generated by the vertices \(\mathbb {V}_i\). Next, we execute the cuPPA algorithm on each of the GPUs with their set of vertices. We can immediately see that computing phase 1 of the algorithm can be done independently in all GPUs. However, resolving the incomplete edges in phase 2 of the algorithm requires careful attention. Due to the nature of the dependency, an incomplete edge on a GPU may require the information resident on the memory of another GPU. Therefore, a synchronization among the GPUs is required. Such a synchronization technique between CPUs with distributed memory was presented in [4]. Although the technique can be adapted for synchronization between the GPUs, it requires complex and intricate communications between the GPUs. NVidia CUDA offers another scheme for accessing memory across multiple GPUs called the “unified memory addressing.” In this case, a single memory address space accessible from any processor in a system is available from the CUDA runtime application programming interface. Therefore, any GPU can access the memory of other GPUs. However, due to the nature of random and sparse memory access, the approach would not yield the desired benefit. In the next section, we present cuPPAHash, an alternative algorithm to generate networks using the preferential attachment model using hash functions instead of pseudorandom number generators.
4.1 cuPPAHash: A Hash FunctionBased Implementation
Notice that the dependency of generating an edge on other vertices only arises while creating a copy edge, i.e., when a vertex u tries to connect to an random end point of another vertex v. We adapt an idea previously used for communicationfree parallel generation of BA graphs [24] to a GPU setting. Consider a vertex u copying an end point from a vertex v. Rather than looking up this value from a memory cell that is filled when vertex v is generated, the end point is recomputed independently. This is possible using a hash function to generate the random numbers instead of using pseudorandom numbers. This approach has the additional benefit that the exact same graph can be reproduced using the same hash functions. We also extend [24] by developing a more general preferential attachmentbased algorithm using the copy model called cuPPAHash.
A simplified pseudocode of cuPPAHash is presented in Algorithm 2. The initial set of vertices is first divided into g mutually exclusive subsets \(\mathbb {V}_1,\ldots ,\mathbb {V}_g\) where g is the number of GPUs. Next each GPU k processes the vertices in set \(\mathbb {V}_k\) using the procedure cuPPAHash (\(\mathbb {V}_k\)). We further partition the set of vertices \(\mathbb {V}_k\) into T subsets \(V_0, V_1, \ldots , V_{T1}\) where T is the number of threads on GPU k. Thread i executes the copy model on the vertices in set \(V_i\) (line 3). For each of the d outgoing edge of a vertex v, the function CalculateEdge calculates the endpoint vertex using the copy model. The \(\ell\)th outgoing edge of the vertex v is uniquely denoted by the index \(e = vd + \ell\) (line 7). The edge index e is used to generate a hash value r using a hash function. We used a 64bit CRC64 as our hash function. Note that r denotes the index of the lth outgoing edge of vertex u calculated in line 9. If \(u<d\) then u denotes an initial vertex and we connect \(F_\ell (v)\) to u (line 11). Otherwise, we compute a floating point number r using a floating point version of the hash function (line 14). If \(r<p\) (i.e., with probability p), we connect \(F_\ell (v)\) to u. Otherwise, we calculate the outgoing edge of \(F_l(u)\) (line 17) recursively. In the actual implementation, we use an iterative function instead of the recursive one.
Note that the algorithm does not require to access any GPU memory pertaining to other GPUs. Instead all copy edges are essentially recomputed. Therefore, this approach requires more computation than the original cuPPA algorithm. However, due to the independent computations, the algorithm scales very well to multiple GPUs as shown in the experimental section.
5 Experimental Results
In this section, we evaluate our algorithm and its performance by experimental analysis. In the following sections, we denote our first algorithm using pseudorandom number generators as cuPPAPure and the second algorithm using hash function as cuPPAHash. We demonstrate the accuracy of our algorithm by showing that our algorithm produces networks with powerlaw degree distribution as desired. We also compare the runtime of our algorithm using other sequential and parallel algorithms.
5.1 Hardware and Software
We used a computer consisting of 6 AMD Phenom(tm) II 6174 processor with 3.3 GHz clock speed and 64 GB system memory. The machine also incorporates a NVidia 1080 GPU with 8 GB memory. The operating system is Ubuntu 16.04 LTS, and all software on this machine was compiled with GNU gcc 4.6.3 with optimization flags O3. The CUDA compilation tools V8 were used for the GPU code along with nvcc compiler. In additional experiments, we used another system consisting of 4 NVidia Tesla P100 GPUs with 16 GB memory each.
5.2 Degree Distribution
5.3 Visualization of Generated Graphs
In order to gain an idea of the structure and degree distributions, we obtained a visualization of some of the networks generated by our algorithm. We generated the visualizations using a popular network visualization tool called Gephi. Bearing aesthetics in mind and to minimize undue clutter, we focused on a few small networks by choosing \(n=10{,}000\), \(p=0.5\), and \(d=1,2,4\). The visualizations are shown in Figs. 4, 5, and 6.
5.4 Effect of Edge Probability on Degree Distribution
5.5 Waiting Queue Size of cuPPAPure
However, as the round progresses, the maximum size of the waiting queue decreases significantly as shown in Fig. 9. For this figure, we also ran cuPPA using 1280 CUDA threads (20 blocks and 64 threads per block) to generate networks with \(d=512, 256, 128, 64\) and \(p=0.5\). Each CUDA thread processes exactly one vertex per round. Only the first 100 rounds are shown for brevity. From Fig. 9, we can see that as the round progresses, the size of the waiting queue per round decreases dramatically for all different values of d. This indicates that we could process more vertices in later rounds using the same amount of queue memory. Therefore, we can dynamically change the size of the segments between two consecutive rounds to increase parallelization. Based on these observations regarding the size of the waiting queue, we designed an adaptive version of cuPPAPure that monitors the maximum size of the waiting queue and manages the segment size accordingly as discussed in Sect. 3.3. We call this version cuPPADynamic and use it for all other experiments.
5.6 Runtime Performance
In this section, we analyze the runtime and performance of cuPPAPure and cuPPAHash relative to other algorithms and show the variation of performances against various parameters.
5.6.1 Runtime Comparison with Existing Algorithms (Single GPU)

Sequential Algorithms: We compare cuPPA with two efficient sequential algorithms: SBA [9] and SCM [17]. For SCM, we used two implementations: one with the pseudorandom number generators (called SCMPure) and the other with hash functions (called SCMHash). We also compared our algorithm with a reference sequential graph generation library from the Graph500 [1] reference code that uses SKG to generate networks.
As shown in Fig. 10, SKG from Graph500 takes the longest time to generate 2B scalefree networks—25.39 minutes. In comparison, our GPUbased algorithm is \(650\,\times\) faster.
We also found that SCMPure is slightly faster than the SBA algorithm. The hashbased SCMHash essentially recomputes all the copy edges and therefore takes approximately 70% more time than the SCMPure algorithm. However, the hashing technique is shown to scale to a large number of processors making it a viable candidate for large network generation using many processors [24]. On the other hand, the GPUbased cuPPAPure generates the network in just 2.32 s on the NVidia 1080 GPU with \(78\,\times\) to \(94\,\times\) speedup. Also note that cuPPAHash is slightly slower than cuPPAPure on a single GPU due to more computation.

Parallel Algorithms: We also compared cuPPA with a distributedmemory (PPADM) [4] and a sharedmemory (PPASM) [7] parallel algorithms. As shown in Fig. 10, both of the cuPPA algorithms outperform PPADM on a system with 24 processors. The main reason is that unlike PPADM, cuPPA algorithms do not require complex synchronizations and message communications.
Due to the unavailability of the PPASM code, we compared the runtime to generate the largest graph (\(n=10^7, d=10\)) reported in [7] with the corresponding runtime of cuPPA. PPASM generates the network using 16 cores of Intel Xeon CPU E52698 2.30 GHz in approximately 7.5 s, whereas cuPPAPure generates the same network in just 0.3 s.
5.6.2 Runtime Versus Number of Vertices (Single GPU)
5.6.3 Runtime Versus Degree of Preferential Attachment (Single GPU)
Next, we examine the runtime performance of cuPPA with increasing d. The runtime is shown in Fig. 13. Here, we set \(n=7812500\), vary \(p=\{0, 0.00001,\) \(0.001, 0.25, 0.5, 0.75, 1\}\), and vary \(d=\{1, 2, 4, 8, 16, 32, 64, 128\}\) to see how the runtime changes for increasing value of d for different p. As seen from the figure, with increasing d, the runtime increases almost linearly. Therefore, the algorithm is observed to scale well for increasing value of d. Note that higher values of d are typically unlikely. However, we included higher values of d for performance measurement purpose. Also notice that the runtime is the largest for \(p=0\). With a small value of \(p=0.00001\), the runtime drops significantly and does not change much for higher values of p. Since the typical values of p are much larger than 0, this observation suggests that cuPPA performs well for realworld scenarios.
5.6.4 Runtime Versus Probability of Copy Edge (Single GPU)
5.6.5 Runtime Varied with the Number of Threads (Single GPU)
5.7 Runtime Performance of cuPPAHash (Multiple GPUs)
Next, we evaluate the performance of cuPPAHash for multiple GPUs. In this experiment, we used a machine consisting of 4 NVidia Tesla P100 GPUs with 16 GB memory each.
5.7.1 Strong Scaling
5.7.2 Generating Large Networks
Using cuPPAHash with 4 GPUs, we are able to generate a network of 16B edges (\(n=2B\) and \(d=8\)) in just 7 s. That represents a rate of 2.29 billion edges per second, which is unprecedented in this domain.
6 Related Work
Although the concepts of random networks have been used and well studied over the last several decades, efficient algorithms to generate the networks were not available until recently. The first efficient sequential algorithm to generate Erdős–Rényi and Barabási–Albert networks was proposed in [9]. A distributedmemorybased algorithm to generate preferential attachment networks was proposed in [26]. However, their algorithm was not exact, rather an approximate algorithm and required manually adjusting several control parameters. The first exact distributedmemorybased parallel algorithm using the copy model was proposed in [4]. Another distributedmemorybased parallel algorithm using the Barabási–Albert model was proposed in [22, 24]. However, instead of using pseudorandom number generators, they used hash functions to generate the networks. A sharedmemorybased parallel algorithm using the copy model was proposed in [7].
Several other theoretical studies were done on the preferential attachmentbased models. Machta and Machta [21] described how an evolving network can be generated in parallel. Dorogovtsev et al. [13] proposed a model that can generate graphs with fattailed degree distributions. In this model, starting with some random graphs, edges are randomly rewired according to some preferential choices. There exists other popular network models to generate networks with powerlaw degree distribution. RMAT [10] and stochastic Kronecker graph (SKG) [19] models can generate networks with powerlaw degree distribution using matrix multiplication. Due to its simpler parallel implementation, the Graph500 group [1] choose the SKG model in their supercomputer benchmark. Highly scalable generators for ErdősRenyi, 2D/3D random geometric graphs, 2D/3D Delaunay graphs, and hyperbolic random graphs are described in [15]. The corresponding software library release also includes an implementation of the algorithm described in [24]. An efficient and scalable algorithmic method to generate Chung–Lu, block twolevel Erdős–Renyi (BTER), and stochastic blockmodels was also presented in [5].
There is a lack of GPUbased network generators in the literature. A GPUbased algorithm for generating Erdős–Rényi networks was presented in [23]. Another GPUbased algorithm for generating networks using the smallworld model [25] was presented in [18]. However, until recently no GPUbased algorithm existed for the preferential attachment model. We introduced cuPPA as the first preferential attachmentbased algorithm on the GPU using the copy model [3].
7 Conclusion
A novel GPUbased algorithm, named cuPPA, has been presented, with a detailed performance study, and its combination of its scale and speed has been tested by achieving the ability to generate networks with up to two billion edges in under 3 s of wall clock time. The algorithm is customizable with respect to the structure of the network by varying a single parameter, namely, a probability measure that captures the preference style of new edges in the preferential attachment model. Also, a high amount of concurrency in the generator’s workload per thread or processor is observed when that probability is at very small fractions greater than zero. In future work, we intend to exploit code profiling tools for further optimization of the runtime and memory usage on the GPU. Also, the algorithm needs to be extended to exploit multiple GPUs that may be colocated within the same node. This would require periodic data synchronization across GPUs, which can be efficiently achieved using the NVidia Collective Communication Library (NCCL). Additional future work involves porting to GPUs spanning multiple nodes, and also hybrid CPUGPU scenarios in order to utilize unused cores of multicore CPUs. Methods to incorporate other network generator models can also be explored with our cuPPA as a starting point. Finally, future work is needed in converting our internal, GPUbased graph representation to other popular network formats for usability.
Footnotes
Notes
Acknowledgements
Funding was provided by Oak Ridge National Laboratory (Grant No. 3X012DCS).
References
 1.(2010) Graph 500. http://www.graph500.org/
 2.Alam M (2016) HPCbased parallel algorithms for generating random networks and some other network analysis problems. Ph.D. thesis, Virginia TechGoogle Scholar
 3.Alam M, Perumalla KS (2017) Gpubased parallel algorithm for generating massive scalefree networks using the preferential attachment model. In: IEEE international conference on Big Data (Big Data), pp 3302–3311. https://doi.org/10.1109/BigData.2017.8258315
 4.Alam M, Khan M, Marathe MV (2013) Distributedmemory parallel algorithms for generating massive scalefree networks using preferential attachment model. In: International conference for high performance computing, networking, storage and analysis. https://doi.org/10.1145/2503210.2503291
 5.Alam M, Khan M, Vullikanti A, Marathe M (2016) An efficient and scalable algorithmic method for generating large: Scale random graphs. In: Proceedings of the international conference for high performance computing, networking, storage and analysis. IEEE Press, Piscataway, NJ, USA, SC ’16, pp 32:1–32:12. http://dl.acm.org/citation.cfm?id=3014904.3014947
 6.Albert R, Jeong H, Barabási AL (2000) Error and attack tolerance of complex networks. Nature. https://doi.org/10.1038/35019019 CrossRefGoogle Scholar
 7.Azadbakht K, Bezirgiannis N, de Boer FS, Aliakbary S (2016) A highlevel and scalable approach for generating scalefree graphs using active objects. In: Proceedings of the annual ACM symposium on applied computingGoogle Scholar
 8.Barabási AL, Albert R (1999) Emergence of scaling in random networks. Science. https://doi.org/10.1126/science.286.5439.509
 9.Batagelj V, Brandes U (2005) Efficient generation of large random networks. Phys Rev E. https://doi.org/10.1103/PhysRevE.71.036113 CrossRefGoogle Scholar
 10.Chakrabarti D, Zhan Y, Faloutsos C (2004) Rmat: A recursive model for graph mining. In: SIAM international conference on data mining, pp 442–446. https://doi.org/10.1137/1.9781611972740.43
 11.Chassin DP, Posse C (2005) Evaluating North American electric grid reliability using the Barabási–Albert network model. Physica A. https://doi.org/10.1016/j.physa.2005.02.051
 12.Dorogovtsev S, Mendes J (2002) Evolution of networks. In: Advances in physics, vol 51. https://doi.org/10.1080/00018730110112519
 13.Dorogovtsev S, Mendes J, Samukhin A (2003) Principles of statistical mechanics of uncorrelated random networks. Nucl Phys B. https://doi.org/10.1016/S05503213(03),005042
 14.Erdős P, Rényi A (1960) On the evolution of random graphs. Publications of the Mathematical Institute of the Hungarian Academy of Sciences, pp 17–61Google Scholar
 15.Funke D, Lamm S, Sanders P, Schulz C, Strash D, von Looz M (2018) Communicationfree massively distributed graph generation. In: 32nd IEEE international parallel & distributed processing symposium (IPDPS), to appear, preprint arXiv:1710.07565
 16.Kleinberg J, Kumar R, Raghavan P, Rajagopalan S, Tomkins A (1999) The Web as a graph: measurements, models, and methods. In: Annual international conference on computing and combinatoricsGoogle Scholar
 17.Kumar R, Raghavan P, Rajagopalan S, Sivakumar D, Tomkins A, Upfal E (2000) Stochastic models for the web graph. In: Annual symposium on foundations of computer science, IEEE Comput. Soc. https://doi.org/10.1109/SFCS.2000.892065
 18.Leist A, Hawick K (2011) Graph generation on GPUs using dynamic memory allocation. In: International conference on parallel and distributed processing techniques and applicationsGoogle Scholar
 19.Leskovec J (2010) Kronecker graphs: an approach to modeling networks. J Mach Learn Res 11:985–1042MathSciNetzbMATHGoogle Scholar
 20.Leskovec J, Horvitz E (2008) Planetaryscale views on a large instantmessaging network. In: International conference on World Wide Web, ACM Press. https://doi.org/10.1145/1367497.1367620
 21.Machta B, Machta J (2005) Parallel dynamics and computational complexity of network growth models. Phys Rev E 71(2):26704. https://doi.org/10.1103/PhysRevE.71.026704 CrossRefGoogle Scholar
 22.Meyer U, Penschuck M (2016) Generating massive scalefree networks under resource constraints. In: Proceeding of the workshop on algorithm engineering and experiments (ALENEX)Google Scholar
 23.Nobari S, Lu X, Karras P, Bressan S (2011) Fast random graph generation. In: International conference on extending database technology, p 331. https://doi.org/10.1145/1951365.1951406
 24.Sanders P, Schulz C (2016) Scalable generation of scalefree graphs. Info Proc LettGoogle Scholar
 25.Watts DJ, Strogatz SH (1998) Collective dynamics of ‘smallworld’ networks. Nature (6684). https://doi.org/10.1038/30918
 26.Yoo A, Henderson K (2010) Parallel generation of massive scalefree graphs. arXiv CoRRGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.