• Journal of KIISE:Computer Systems and Theory
• /
• v.26 no.9
• /
• pp.1073-1082
• /
• 1999

본 논문에서는 시스템 내의 프로세서들을 효과적으로 사용하기 위한 적응적 프로세서 할당 정책을 제안한다. 프로그램의 병렬성을 향상시키기 위하여 일반적으로 병렬 처리에 사용될 프로세서 개수를 증가시킨다. 그러나 증가된 프로세서들은 그레인 크기에 변화를 일으키며 이는 캐쉬 성능에 영향을 미친다. 특히 대역이 제한된 공유 버스를 사용하는 시스템에서는 프로세서 개수의 증가는 공유 버스에 대한 접근 경쟁을 크게 증가하므로 버스에서 대기하는 시간이 프로세서 증가에 의한 계산 능력 이득을 상쇄시키는 주요한 원인이 되고 있다. 본 논문에서 제안한 적응적 프로세서 할당 정책은 프로그램이 수행되는 도중에 임의의 기간동안 공유버스에 대기중인 프로세서 분포에 관한 정보를 얻는다. 그리고 이 정보를 바탕으로 프로세서 개수를 변경하는 방법이다. 모의 시험에서 적응적 프로세서 할당 정책은 프로그램들의 버스 트래픽 특성에 따른 최적의 적합한 프로세서 개수를 발견함을 보인다. 그리고 적응적 프로세서 할당 정책은 고정된 프로세서 개수를 사용한 가장 좋은 성능보다는 다소 떨어진 성능을 나타내었으나 시스템의 프로세서 활용성을 높여 효과적 시스템 사용에 기여함을 보인다. Abstract In this paper, the adaptive processor allocation policy is suggested to make effective use of processors in system. To enhance the parallelism, the number of processors used in the parallel computing may be increased. However, increasing the number of processors affects the grain size of the parallel program. Therefore, it affects the cache performance. In particular, when the shared bus is employed, since increasing the number of processors can result in a significant amount of contention to achieve the shared-bus, the increased computing power is offset by the bus waiting time due to these contentions. The adaptive processor allocation policy acquires the information about the distribution of waiting processors on shared bus for any execution period of programs. And it changes the number of processors working in parallel processing during the program's run. Our simulation results show that the adaptive processor allocation policy finds the optimum feasible number of processors based on the bus traffic characteristic of programs. Thus, it contributes to effective system utilization, even though it performs slightly less efficiently than using a fixed number of processors with the best performance.

• Journal of KIISE:Computer Systems and Theory
• /
• v.37 no.2
• /
• pp.76-89
• /
• 2010

The performance of Distributed Heterogeneous Computing System depends on the algorithm which schedules input DAG graph. Among various scheduling algorithms, list scheduling algorithm provides superior performance with low complexity. List scheduling consists of task prioritizing phase and processor allocation phase, but most studies only focus on task prioritizing phase. In this paper, we propose LIP policy which has the same complexity with traditional allocation policies but has superior performance. The performance of LIP has been observed by applying them to task prioritizing phase of traditional list scheduling algorithms, HCPT, HEFT, GCA, and PETS. The results show that LIP has better performance than insertion-based policy and non-insertion-based policy, which are traditional processor allocation policies.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2010.11a
• /
• pp.1818-1821
• /
• 2010

본 논문에서는 Xen 가상 머신에서 연성 실시간 처리를 제공하기 위하여 하이퍼바이저(Hypervisor)의 VCPU 할당 정책을 제시한다. 현재 Xen 에서는 가상의 CPU (VCPU)를 스케줄링하여 전체 시스템을 관리한다. 본 연구에서는 실시간 태스크를 지원하는 VCPU 할당 정책의 사전 작업으로 기존의 Xen VCPU 할당 정책을 분석한다. 분석된 할당 정책의 단점으로 멀티프로세서 환경에서의 불필요한 VCPU 마이그레이션(Migration) 으로 인한 오버헤드를 확인하고 기존 정책의 단점을 보완하기 위한 새로운 할당 정책을 제시하여 Xen 플랫폼에서 연성 실시간을 지원하는 VCPU 할당 정책을 제안하고 있다.

• The KIPS Transactions:PartA
• /
• v.11A no.3
• /
• pp.199-206
• /
• 2004

Under the multiprogrammed situation, the performance of multiprocessor system is affected by the process allocation policy of the operating systems. The lowest communication cost can be achieved when the related processes positioned to the adjacent processors. While the effective allocation is quite difficult to the real situation, and the processing of the allocation policy consumes some computation time. The dual-ring CC-NUMA systems exhibit a quite performance difference according to the process a1location policy due to a lot of unbalanced memory transactions on the interconnection networks. In this paper, I propose a load balanced dual-link CC-NUMA system that does not requires the processes allocation policy. By the program-driven simulation results. the proposed system shows no remarkable difference according to the allocation policy while the dual-ring systems shows 10％ performance improvement by the process allocation. In addition, the proposed system outperforms the dual~ring systems about 1.5 times.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2011.11a
• /
• pp.141-144
• /
• 2011

최근 이기종 멀티프로세서 시스템에서의 병렬화를 위해 범용 CPU 와 다른 컴퓨팅 장치들간의 다양한 연동 기술들이 부각되고 있다. 멀티프로세서 프로그래밍 모델인 OpenMP 는 가장 널리 사용되는 병렬 프로그래밍 언어이지만 기존 OpenMP 의 작업 할당 정책으로는 프로세서간 로드밸런싱을 문제를 해결할 수 없다는 한계점을 가지고 있다. 본 논문에서는 기존 OpenMP 의 작업할당 문제를 해결할 수 있는 알고리즘을 제안한다. 제안하는 알고리즘은 SMP(Symmetric Multi Processing) 구조뿐만 아니라 AMP(명령어 구조는 같으나 동작 속도가 다른 이질 멀티프로세서 구조)에서도 작업부하균형을 효과적으로 실행할 수 있다.

• Proceedings of the Korean Information Science Society Conference
• /
• 2000.04a
• /
• pp.86-88
• /
• 2000

멀티미디어 태스크는 기존 범용 운영 체제의 시분할 스케줄러에서 만족시킬 수 없는 시간적 요구사항을 가진다. 이런 태스크를 기존의 시분할 태스크와 함께 서비스하기 위해서는 새로운 스케줄링 프레임워크가 필요하다. FQ(Fair Queueing)은 태스크의 공유비율에 비례하여 자원을 할당하는 방법으로 이질적인 태스크(멀티미디어 태스크, 일반 시분할 태스크)가 공존하는 개방적인 환경에서의 스케줄링 정책으로 적합하다는 특징이 있다. 본 논문에서는 FQ의 종류인 WFQ(Weighted Fair Queueing)를 개선하여 하나의 스케줄러에서 다른 두 부류의 태스크를 모두 처리하는 요구비율 기반의 프로세서 예약 기법을 제안한다. 실시간 태스크와 시분할 태스크를 처리하기 위해서 실시간 부류의 태스크를 우선적으로 배치하고, 실시간 부류 태스크의 실행 사이에 시분할 태스크를 스케줄하여 실시간 태스크에 대해서 보장된 서비스를, 시분할 태스크에 대해서는 이 태스크에 할당된 예약만큼의 프로세서 시간을 제공한다. 모의 실험에서는 제안한 프로세서 예약 방식이 실시간 태스크와 시분할 태스크를 효율적으로 처리하며 기존의 WFQ보다 더 안정적임을 보인다.

• The Transactions of the Korea Information Processing Society
• /
• v.1 no.3
• /
• pp.367-379
• /
• 1994

This paper describes scheduling algorithms of the UNIX operating system and shows an analytical approach to approximate the average conditional response time for a process in the UNIX operating system. The average conditional response time is the average time between the submittal of a process requiring a certain amount of the CPU time and the completion of the process. The process scheduling algorithms in thr UNIX system are based on the priority service disciplines. That is, the behavior of a process is governed by the UNIX process schuduling algorithms that (ⅰ) the time-shared computer usage is obtained by allotting each request a quantum until it completes its required CPU time, (ⅱ) the nonpreemptive switching in system mode and the preemptive switching in user mode are applied to determine the quantum, (ⅲ) the first-come-first-serve discipline is applied within the same priority level, and (ⅳ) after completing an allotted quantum the process is placed at the end of either the runnable queue corresponding to its priority or the disk queue where it sleeps. These process scheduling algorithms create the round-robin effect in user mode. Using the round-robin effect and the preemptive switching, we approximate a process delay in user mode. Using the nonpreemptive switching, we approximate a process delay in system mode. We also consider a process delay due to the disk input and output operations. The average conditional response time is then obtained by approximating the total process delay. The results show an excellent response time for the processes requiring system time at the expense of the processes requiring user time.

• Journal of KIISE:Computer Systems and Theory
• /
• v.27 no.7
• /
• pp.665-674
• /
• 2000

In data parallel programs incurring high cache locality, the choice of grain sizes affects cache performance. Though the grain sizes chosen provide fair load balance among processors, the grain sizes that ignore underlying caching effect result in address interferences between grains allocated to a processor. These address interferences appear to have a negative impact on the cache locality, since they result in cache conflict misses. To address this problem, we propose a best grain size driven from a cache size and the number of processors based on direct mapped cache's characteristic. Since the proposed method does not map the grains to the same location in the cache, cache conflict misses are reduced. Simulation results show that the proposed best grain size substantially improves the performance of tested data parallel programs through the reduction of cache misses on direct-mapped caches.

• Journal of KIISE:Computer Systems and Theory
• /
• v.34 no.7
• /
• pp.304-318
• /
• 2007

Grid computing is a service that share geographically distributed computing resources through high speed network. In this paper, we propose hybrid scheduling scheme which considers not only meta-scheduling scheme to distribute the job between the nodes of grid computing system but also the job scheduling to distribute the job within the local nodes. According to the number of processors needed and expected execution time, the job with high priority is allocated to job queue while the one with low priority and remote job are allocated to backfill queue. We evaluate the proposing scheme through the various experiments and the results show that the utilization of grid computing system increases and the job slowdown decreases.

• The KIPS Transactions:PartA
• /
• v.19A no.3
• /
• pp.129-138
• /
• 2012

In distributed memory architectures like clusters, each node stores a portion of data. How data is distributed across nodes influences the performance of such systems. The data distribution scheme is the strategy to distribute data across nodes and realize parallel data processing. Due to various reasons such as maintenance, scale up, upgrade, etc., the performance of nodes in a cluster can often become non-identical. In such clusters, data distribution without considering performance cannot efficiently distribute data on nodes. In this paper, we propose a new data distribution scheme based on the number of cores in nodes. We use the number of cores as the performance factor. In our data distribution scheme, each node is allocated an amount of data proportional to the number of cores in it. We implement our data distribution scheme using the Chapel language. To show our data distribution is effective in reducing the execution time of parallel applications, we implement Mandelbrot Set and ${\pi}$-Calculation programs with our data distribution scheme, and compare the execution times on a cluster. Based on experimental results on clusters of 8-core and 16-core nodes, we demonstrate that data distribution based on the number of cores can contribute to a reduction in the execution times of parallel programs on clusters.