Browsing by Author "Ozdal, M. M."

Now showing 1 - 3 of 3

Open Access
Graph analytics accelerators for cognitive systems
(Institute of Electrical and Electronics Engineers, 2017) Ozdal, M. M.; Yesil, S.; Kim, T.; Ayupov, A.; Greth, J.; Burns, S.; Ozturk, O.
Hardware accelerators are known to be performance and power efficient. This article focuses on accelerator design for graph analytics applications, which are commonly used kernels for cognitive systems. The authors propose a templatized architecture that is specifically optimized for vertex-centric graph applications with irregular memory access patterns, asynchronous execution, and asymmetric convergence. The proposed architecture addresses the limitations of existing CPU and GPU systems while providing a customizable template. The authors' experiments show that the generated accelerators can outperform a high-end CPU system with up to 3 times better performance and 65 times better power efficiency. © 1981-2012 IEEE.
Open Access
Hypergraph models and algorithms for data-pattern-based clustering
(Springer, 2004) Ozdal, M. M.; Aykanat, Cevdet
In traditional approaches for clustering market basket type data, relations among transactions are modeled according to the items occurring in these transactions. However, an individual item might induce different relations in different contexts. Since such contexts might be captured by interesting patterns in the overall data, we represent each transaction as a set of patterns through modifying the conventional pattern semantics. By clustering the patterns in the dataset, we infer a clustering of the transactions represented this way. For this, we propose a novel hypergraph model to represent the relations among the patterns. Instead of a local measure that depends only on common items among patterns, we propose a global measure that is based on the cooccurences of these patterns in the overall data. The success of existing hypergraph partitioning based algorithms in other domains depends on sparsity of the hypergraph and explicit objective metrics. For this, we propose a two-phase clustering approach for the above hypergraph, which is expected to be dense. In the first phase, the vertices of the hypergraph are merged in a multilevel algorithm to obtain large number of high quality clusters. Here, we propose new quality metrics for merging decisions in hypergraph clustering specifically for this domain. In order to enable the use of existing metrics in the second phase, we introduce a vertex-to-cluster affinity concept to devise a method for constructing a sparse hypergraph based on the obtained clustering. The experiments we have performed show the effectiveness of the proposed framework.
Open Access
A novel method for scaling iterative solvers: avoiding latency overhead of parallel sparse-matrix vector multiplies
(Institute of Electrical and Electronics Engineers, 2015) Selvitopi, R. O.; Ozdal, M. M.; Aykanat, Cevdet
In parallel linear iterative solvers, sparse matrix vector multiplication (SpMxV) incurs irregular point-to-point (P2P) communications, whereas inner product computations incur regular collective communications. These P2P communications cause an additional synchronization point with relatively high message latency costs due to small message sizes. In these solvers, each SpMxV is usually followed by an inner product computation that involves the output vector of SpMxV. Here, we exploit this property to propose a novel parallelization method that avoids the latency costs and synchronization overhead of P2P communications. Our method involves a computational and a communication rearrangement scheme. The computational rearrangement provides an alternative method for forming input vector of SpMxV and allows P2P and collective communications to be performed in a single phase. The communication rearrangement realizes this opportunity by embedding P2P communications into global collective communication operations. The proposed method grants a certain value on the maximum number of messages communicated regardless of the sparsity pattern of the matrix. The downside, however, is the increased message volume and the negligible redundant computation. We favor reducing the message latency costs at the expense of increasing message volume. Yet, we propose two iterative-improvement-based heuristics to alleviate the increase in the volume through one-to-one task-to-processor mapping. Our experiments on two supercomputers, Cray XE6 and IBM BlueGene/Q, up to 2,048 processors show that the proposed parallelization method exhibits superior scalable performance compared to the conventional parallelization method.