• Journal of Korean Society of Industrial and Systems Engineering
• /
• v.38 no.3
• /
• pp.56-63
• /
• 2015

Several fields of science have demanded large-scale workflow support, which requires thousands of CPU cores or more. In order to support such large-scale scientific workflows, high capacity parallel systems such as supercomputers are widely used. In order to increase the utilization of these systems, most schedulers use backfilling policy: Small jobs are moved ahead to fill in holes in the schedule when large jobs do not delay. Since an estimate of the runtime is necessary for backfilling, most parallel systems use user's estimated runtime. However, it is found to be extremely inaccurate because users overestimate their jobs. Therefore, in this paper, we propose a novel system for the runtime prediction based on workload-aware clustering with the goal of improving prediction performance. The proposed method for runtime prediction of parallel applications consists of three main phases. First, a feature selection based on factor analysis is performed to identify important input features. Then, it performs a clustering analysis of history data based on self-organizing map which is followed by hierarchical clustering for finding the clustering boundaries from the weight vectors. Finally, prediction models are constructed using support vector regression with the clustered workload data. Multiple prediction models for each clustered data pattern can reduce the error rate compared with a single model for the whole data pattern. In the experiments, we use workload logs on parallel systems (i.e., iPSC, LANL-CM5, SDSC-Par95, SDSC-Par96, and CTC-SP2) to evaluate the effectiveness of our approach. Comparing with other techniques, experimental results show that the proposed method improves the accuracy up to 69.08%.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.19 no.1
• /
• pp.111-116
• /
• 2019

Recently, GPU expands application domains from graphic processing to various kinds of parallel workloads. However, current GPU systems focus on the maximization of each workload's parallelism through simplified control rather than considering various workload characteristics. This paper classifies the resource usage characteristics of GPU workloads into computing-bound, memory-bound, and dependency-latency-bound, and quantifies the fine-grained bottleneck for efficient workload allocation. For example, we identify the exact bottleneck resources such as single function unit, double function unit, or special function unit even for the same computing-bound workloads. Our analysis implies that workloads can be allocated together if fine-grained bottleneck resources are different even for the same computing-bound workloads, which can eventually contribute to efficient workload allocation in GPU.

• Journal of the Institute of Convergence Signal Processing
• /
• v.5 no.3
• /
• pp.236-242
• /
• 2004

In this paper, we propose an approach to parallelize a load balancing algorithm that was shown to be very effective in distributing workload for parallel computations. Load balancing algorithms are required in executing parallel program efficiently As a parallel computation model, we used dynamically growing tree structure that can be found in many application problems. The load balancing algorithm tries to balance the workload among processors while keeping the communication cost under certain limit. We show how the load balancing algorithm is effectively parallelized on mesh and hypercube interconnection networks, and analyzed the time complexity for each case to show that parallel algorithm actually reduced the various overhead.

• Journal of the Korea Society of Computer and Information
• /
• v.13 no.1
• /
• pp.109-116
• /
• 2008

Even though many studies on parallel rendering based on PC clusters have been done. most of those did not cope with non-uniform scenes, where locations of 3D models are biased. In this work. we have built a PC cluster system with POV-Ray, a free rendering software on the public domain, and developed an adaptive load balancing scheme to optimize the parallel efficiency Especially, we noticed that a frame of 3D animation are closely coherent with adjacent frames. and thus we could estimate distribution of computation amount, based on the computation time of previous frame. The experimental results with 2 real animation data show that the proposed scheme reduces by 40% of execution time compared to the simple static partitioning scheme.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.44 no.4 s.316
• /
• pp.101-108
• /
• 2007

Parallel processing algorithms, coupled with advanced networking and distributed computing architectures, improve the overall computational performance, dependability, and versatility of a digital signal processing system In this paper, novel parallel algorithms are introduced and investigated for advanced sonar algorithm, conventional matched-field processing (CMFP). Based on a specific domain, each parallel algorithm decomposes the sequential workload in order to obtain scalable parallel speedup. Depending on the processing requirement of the algorithm, the computational performance of the parallel algorithm reveals different characteristics. The high-complexity algorithm, CMFP shows scalable parallel performance on the array of DSP processors. The impact on parallel performance due to workload balancing, communication scheme, algorithm complexity, processor speed, network performance, and testbed configuration is explored.

• The Transactions of the Korea Information Processing Society
• /
• v.7 no.9
• /
• pp.2827-2835
• /
• 2000

In recent years, much attention has been focused on improving I/O devices' processing speed which is essential in such large data processing areas as multimedia data processing. And studies on high-performance parallel file systems are considered to be one of such efforts. In this paper, an efficient disk allocation scheme is proposed for high-performance parallel file systems. In other words, the concept of a parallel disk file's parallelism is defined using data declustering characteristic of a given parallel file. With the concept, an efficient disk allocation scheme is proposed which calculates the appropriate degree of data declustering on disks for each parallel file in order to obtain the maximum throughput when more than one parallel file is used at the same time. Since, calculation for obtaining the maximum throughput is too complex as the number of parallel files increases, an approximate disk allocation algorithm is also proposed in this paper. The approximate algorithm is very simple and especially provides very good results when I/O workload is high. In addition, it has shown that the approximate algorithm provides the optimal disk allocation for the maximum throughput when the arrival rate of I/O requests is infinite.

• Journal of applied mathematics & informatics
• /
• v.7 no.3
• /
• pp.1045-1058
• /
• 2000

In usual synchronous parallel computing, workload balance is a crucial factor to reduce idle times of some processors that have finished their jobs earlier than others. However, it is difficult to achieve on a heterogeneous workstation clusters where the available computing power of each processor is unpredictable. As a way to overcome such a problem, the idea of asynchronous methods has grown out and is being increasingly used and studied, but there is none for eigenvalue problems yet. In this paper, we suggest a new asynchronous method to solve some singular matrix problems, that can also be used for finding a certain eigenvector of some matrices.

• ETRI Journal
• /
• v.20 no.4
• /
• pp.301-326
• /
• 1998

In this paper, we have modeled and evaluated a new parallel processing system called Scalable Parallel computer Architecture based on Xbar (SPAX) for commercial applications. SMP systems are widely used as servers for commercial applications; however, they have very limited scalability. SPAX cost-effectively overcomes the SMP limitation by providing both scalability and application portability. To investigate whether the new architecture satisfies the requirements of commercial applications, we have built a system model and a workload model. The results of the simulation study show that the I/O subsystem becomes the major bottleneck. We found that SPAX can still meet the I/O requirement of the OLTP workload as it supports flexible I/O subsystem. We also investigated what will be the next most important bottleneck in SPAX and how to remove it. We found that the newly developed system network called Xcent-Net will not be a bottleneck in the I/O data path. We also show the optimal configuration that is to be considered for system tuning.

• Journal of KIISE:Computing Practices and Letters
• /
• v.13 no.6
• /
• pp.358-370
• /
• 2007

Software simulations have been widely used for the design and application development of a wireless sensor network that is an infrastructure of ubiquitous computing. In this study, we develop a sensor network simulator that can verify the behavior of sensor network applications, estimate execution time and power consumption, and simulate a large-scale sensor network. To implement the simulator, we use an instruction-level parallel discrete-event simulation method. Instruction-level simulation uses executable images loaded into a real sensor board as workload, such that it results in the high degree of details. Parallel simulation makes simulation of a large-scale sensor network possible by distributing workload into multiple computers. The simulator can predict the amount of power consumption based on operating time of modules in a sensor node and counting the number of executed instructions by kind. Also it can simulate ubiquitous applications with various scenarios and debug programs. Instruction traces used as workload for simulations are executable images produced by the cross-compiler for ATmega128L microcontroller.

• Journal of Broadcast Engineering
• /
• v.17 no.5
• /
• pp.742-755
• /
• 2012

In this paper, we propose a parallel deblocking algorithm to resolve workload imbalance when the deblocking filter of high efficiency video coding (HEVC) decoder is parallelized. In HEVC, the deblocking filter which is one of the in-loop filters conducts two-step filtering on vertical edges first and horizontal edges later. The deblocking filtering can be conducted with high-speed through data-level parallelism because there is no dependency between adjacent edges for deblocking filtering processes. However, workloads would be imbalanced among regions even though the same amount of data for each region is allocated, which causes performance loss of decoder parallelization. In this paper, we solve the problem for workload imbalance by predicting the complexity of deblocking filtering with coding unit (CU) depth information at a coding tree block (CTB) and by allocating the same amount of workload to each core. Experimental results show that the proposed method achieves average time saving (ATS) by 64.3%, compared to single core-based deblocking filtering and also achieves ATS by 6.7% on average and 13.5% on maximum, compared to the conventional uniform data-level parallelism.