• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.10
• /
• pp.333-344
• /
• 2022

In a container-based cloud environment, multiple containers can share a graphical processing unit (GPU), and GPU sharing can minimize idle time of GPU resources and improve resource utilization. However, in a cloud environment, GPUs, unlike CPU or memory, cannot logically multiplex computing resources to provide users with some of the resources in an isolated form. In addition, containers occupy GPU resources only when performing GPU operations, and resource usage is also unknown because the timing or size of each container's GPU operations is not known in advance. Containers unrestricted use of GPU resources at any given point in time makes managing resource contention very difficult owing to where multiple containers run GPU tasks simultaneously, and GPU tasks are handled in black box form inside the GPU. In this paper, we propose a container management technique to prevent performance degradation caused by resource competition when multiple containers execute GPU tasks simultaneously. Also, this paper demonstrates the efficiency of container management techniques that analyze and propose the problem of degradation due to resource competition when multiple containers execute GPU tasks simultaneously through experiments.

• KNOM Review
• /
• v.23 no.1
• /
• pp.43-50
• /
• 2020

As General Purpose Graphics Processing Unit (GPGPU) recently plays an essential role in high-performance computing, several cloud service providers offer GPU service. Most cluster orchestration platforms in a cloud environment using containers allocate the integer number of GPU to jobs and do not allow a node shared with other jobs. In this case, resource utilization of a GPU node might be low if a job does not intensively require either many cores or large size of memory in GPU. GPU virtualization brings opportunities to realize kernel concurrency and share resources. However, performance may vary depending on characteristics of applications running concurrently and interference among them due to resource contention on a node. This paper proposes GPU container co-execution framework with multiple server creation and execution based on Kubernetes, container orchestration platform for measuring interference which may be occurred by sharing GPU resources. Performance changes according to scheduling policies were investigated by executing several jobs on GPU. The result shows that optimal scheduling is not possible only considering GPU memory and computing resource usage. Interference caused by co-execution among applications is measured using the framework.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2021.07a
• /
• pp.541-542
• /
• 2021

클라우드 컴퓨팅 환경에서는 고성능 컴퓨팅을 지원하기 위해 사용자에게 GPU(Graphic Processing Unit)가 할당된 가상머신을 제공하여 사용자가 고성능 응용을 실행할 수 있도록 지원한다. 일반적인 컴퓨팅 환경에서 한 명의 사용자가 GPU를 독점해서 사용하기 때문에 자원 경쟁으로 인한 문제가 상대적으로 적게 발생하지만 독립적인 여러 사용자가 컴퓨팅 자원을 공유하는 클라우드 환경에서는 자원 경쟁으로 인해 서로 성능 영향을 미치는 문제를 발생시킨다. 본 논문에서는 여러 개의 가상머신이 단일 GPU를 공유하는 RPC(Remote Procedure Call) 기반 GPU 가상화 환경에서 다수의 가상머신이 GPGPU(General Purpose computing on Graphics Processing Units) 작업을 수행할 때 GPU 메모리 입력 경쟁으로 인해 발생하는 커널 함수의 실행 지연 문제를 분석한다.

• Journal of Internet Computing and Services
• /
• v.14 no.6
• /
• pp.41-47
• /
• 2013

GP-GPUs are general purposed GPUs for numerical computation based on multiple threads which are originally for graphic processing. GP-GPUs provide cache memory in a form of shared memory which user programs can access directly, unlikely typical cache memory. In this research, we implemented the parallel block LU decomposition program to utilize cache memory in GP-GPUs. The parallel blocked LU decomposition program designed with Nvidia CUDA C run 7~8 times faster than nun-blocked LU decomposition program in the same GP-GPU computation environment.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.50 no.4
• /
• pp.108-116
• /
• 2013

Many algorithms have been introduced to improve performance by using massively parallel systems, which consist of several hundreds of processors. A typical example is a GPU system of many processors which uses shared memory. In the case of image filtering algorithms, which make references to neighboring points, the shared memory helps improve performance by frequently accessing adjacent pixels. However, using shared memory requires rewriting the existing codes and consequently results in complexity of the codes. Recent GPU systems support both L1 and L2 cache along with shared memory. Since the L1 cache memory is located in the same area as the shared memory, the improvement of performance is predictable by using the cache memory. In this paper, the performance of cache and shared memory were compared. In conclusion, the performance of cache-based algorithm is very similar to the one of shared memory. The complexity of the code appearing in a shared memory system, however, is resolved with the cache-based algorithm.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.37-39
• /
• 2022

KISTI(한국과학기술정보연구원)는 슈퍼컴퓨터 5호기 메인시스템인 Nurion과 보조시스템인 Neuron을 연구자들에게 서비스하고 있다. Neuron은 메인시스템인 Nurion이 Intel Knights Landing 프로세서가 장착된 클러스터로 구성됨에 따라 인공지능, 빅데이터에 관한 연구 인프라 수요를 충족시키기 위해 GPU를 장착한 이기종 클러스터로 구성되어 있다. Neuron은 연구자들에게 효율적으로 계산 자원을 배분하기 위해 SLURM 작업배치스케줄러의 공유노드 정책을 이용하여 한 개의 계산노드에서 다수 개의 작업이 수행될 수 있는 환경으로 서비스되고 있다. 본 논문에서는 공유노드 정책으로 운영 중인 클러스터 시스템에서 작업별로 GPU 사용 통계 데이터를 생성하는 기법을 소개한다.

• KIISE Transactions on Computing Practices
• /
• v.20 no.9
• /
• pp.507-512
• /
• 2014

Recently, Distributed computing processing begins using both CPU(Central processing unit) and GPU(Graphic processing unit) to improve the performance to overcome darksilicon problem which cannot use all of the transistors because of the electric power limitation. There is an integrated graphics processor that CPU and GPU share memory and Last level cache(LLC). But, There is no LLC access rules between CPU and GPU, so if GPU and CPU processes run together at the same time, performance of both processes gets worse because of the contention on the LLC. This Paper gives evidence to prove the need of the Cache Partitioning and is mentioned about the cache partitioning design using page coloring to allocate the L3 Cache space only for the GPU process to guarantee GPU process performance.

• The Journal of Korean Institute of Next Generation Computing
• /
• v.15 no.2
• /
• pp.39-49
• /
• 2019

Recently, CPU-GPU integrated heterogeneous multicore processors have been widely used for improving the performance of computing systems. Heterogeneous multicore processors integrate CPUs and GPUs on a single chip where CPUs and GPUs share the LLC(Last Level Cache). This causes a serious cache contention problem inside the processor, resulting in significant performance degradation. In this paper, we propose the partitioned LLC architecture to solve the cache contention problem in heterogeneous multicore processors. We analyze the performance impact varying the LLC size of CPUs and GPUs, respectively. According to our simulation results, the bigger the LLC size of the CPU, the CPU performance improves by up to 21%. However, the GPU shows negligible performance difference when the assigned LLC size increases. In other words, the GPU is less likely to lose the performance when the LLC size decreases. Because the performance degradation due to the LLC size reduction in GPU is much smaller than the performance improvement due to the increase of the LLC size of the CPU, the overall performance of heterogeneous multicore processors is expected to be improved by applying partitioned LLC to CPUs and GPUs. In addition, if we develop a memory management technique that can maximize the performance of each core in the future, we can greatly improve the performance of heterogeneous multicore processors.

• KIPS Transactions on Computer and Communication Systems
• /
• v.4 no.7
• /
• pp.213-218
• /
• 2015

Given two strings X and Y (|X|=m, |Y|=n) over an alphabet ${\Sigma}$, the extended edit distance between X and Y can be computed using dynamic programming in O(mn) time and space. Recently, a parallel algorithm that takes O(m+n) time and O(mn) space using m threads to compute the extended edit distance between X and Y was presented. In this paper, we present an improved parallel algorithm using the shared memory on GPU. The experimental results show that our parallel algorithm runs about 19~25 times faster than the previous parallel algorithm.

• KIPS Transactions on Computer and Communication Systems
• /
• v.10 no.5
• /
• pp.123-136
• /
• 2021

RPC (Remote Procedure Call)-based Graphics Processing Unit (GPU) virtualization technology is one of the technologies for sharing GPUs with multiple user virtual machines. However, in a cloud environment, unlike CPU or memory, general GPUs do not provide a resource isolation technology that can limit the resource usage of virtual machines. In particular, in an RPC-based virtualization environment, since GPU tasks executed in each virtual machine are performed in the form of multi-process, the lack of resource isolation technology causes performance degradation due to resource competition. In addition, the GPU memory competition accelerates the performance degradation as the resource demand of the virtual machines increases, and the fairness decreases because it cannot guarantee equal performance between virtual machines. This paper, in the RPC-based GPU virtualization environment, analyzes the performance degradation problem caused by resource contention when the GPU memory requirement of virtual machines exceeds the available GPU memory capacity and proposes a GPU memory management technique to solve this problem. Also, experiments show that the GPU memory management technique proposed in this paper can improve the performance of GPGPU tasks.