PARMA-DITAM@HiPEAC最新文献

筛选

英文中文

COLA-Gen: Active Learning Techniques for Automatic Code Generation of Benchmarks COLA-Gen:自动生成基准代码的主动学习技术

PARMA-DITAM@HiPEAC Pub Date : 1900-01-01 DOI: 10.4230/OASIcs.PARMA-DITAM.2022.3

M. Berezov, Corinne Ancourt, Justyna Zawalska, Maryna Savchenko

{"title":"COLA-Gen: Active Learning Techniques for Automatic Code Generation of Benchmarks","authors":"M. Berezov, Corinne Ancourt, Justyna Zawalska, Maryna Savchenko","doi":"10.4230/OASIcs.PARMA-DITAM.2022.3","DOIUrl":"https://doi.org/10.4230/OASIcs.PARMA-DITAM.2022.3","url":null,"abstract":"Benchmarking is crucial in code optimization. It is required to have a set of programs that we consider representative to validate optimization techniques or evaluate predictive performance models. However, there is a shortage of available benchmarks for code optimization, more pronounced when using machine learning techniques. The problem lies in the number of programs for testing because these techniques are sensitive to the quality and quantity of data used for training. Our work aims to address these limitations. We present a methodology to efficiently generate benchmarks for the code optimization domain. It includes an automatic code generator, an associated DSL handling, the high-level specification of the desired code, and a smart strategy for extending the benchmark as needed. The strategy is based on Active Learning techniques and helps to generate the most representative data for our benchmark. We observed that Machine Learning models trained on our benchmark produce better quality predictions and converge faster. The optimization based on the Active Learning method achieved up to 15% more speed-up than the passive learning method using the same amount of data. The experiments were run on Intel® Core™ i7-8650U 4C/4T @1.90GHz with capacity caches of L1: 32KB, L2: 256KB, L3: 8192KB and 32GB DDR4 DIMM RAM, Phys. cores: 4, Compiler: GCC 5.4.0, Number of Threads: 4, Opt. level: -O3","PeriodicalId":436349,"journal":{"name":"PARMA-DITAM@HiPEAC","volume":"11 suppl_1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1900-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133712538","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Resource Aware GPU Scheduling in Kubernetes Infrastructure Kubernetes基础设施中的资源感知GPU调度

PARMA-DITAM@HiPEAC Pub Date : 1900-01-01 DOI: 10.4230/OASIcs.PARMA-DITAM.2021.4

Aggelos Ferikoglou, Dimosthenis Masouros, Achilleas Tzenetopoulos, S. Xydis, D. Soudris

{"title":"Resource Aware GPU Scheduling in Kubernetes Infrastructure","authors":"Aggelos Ferikoglou, Dimosthenis Masouros, Achilleas Tzenetopoulos, S. Xydis, D. Soudris","doi":"10.4230/OASIcs.PARMA-DITAM.2021.4","DOIUrl":"https://doi.org/10.4230/OASIcs.PARMA-DITAM.2021.4","url":null,"abstract":"Nowadays, there is an ever-increasing number of artificial intelligence inference workloads pushed and executed on the cloud. To effectively serve and manage the computational demands, data center operators have provisioned their infrastructures with accelerators. Specifically for GPUs, support for efficient management lacks, as state-of-the-art schedulers and orchestrators, threat GPUs only as typical compute resources ignoring their unique characteristics and application properties. This phenomenon combined with the GPU over-provisioning problem leads to severe resource under-utilization. Even though prior work has addressed this problem by colocating applications into a single accelerator device, its resource agnostic nature does not manage to face the resource under-utilization and quality of service violations especially for latency critical applications. In this paper, we design a resource aware GPU scheduling framework, able to efficiently colocate applications on the same GPU accelerator card. We integrate our solution with Kubernetes, one of the most widely used cloud orchestration frameworks. We show that our scheduler can achieve 58.8% lower end-to-end job execution time 99%-ile, while delivering 52.5% higher GPU memory usage, 105.9% higher GPU utilization percentage on average and 44.4% lower energy consumption on average, compared to the state-of-the-art schedulers, for a variety of ML representative workloads. 2012 ACM Subject Classification Computing methodologies; Computer systems organization → Cloud computing; Computer systems organization → Heterogeneous (hybrid) systems; Hardware → Emerging architectures","PeriodicalId":436349,"journal":{"name":"PARMA-DITAM@HiPEAC","volume":"14 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1900-01-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130258507","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 1

首页上一页