Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00081
Cleverton Vicentini, A. Santin, E. Viegas, Vilmar Abreu
Cloud computing is intrinsically based on multi-tenancy, which enables a physical host to be shared amongst several tenants (customers). In this context, for several reasons, a cloud provider may overload the physical machine by hosting more tenants that it can adequately handle. In such a case, a tenant may experience application performance issues. However, the tenant is not able to identify the causes, since most cloud providers do not provide performance metrics for customer monitoring, or when they do, the metrics can be biased. This study proposes a two-tier auditing model for the identification of multi-tenancy issues within the tenant domain. Our proposal relies on machine learning techniques fed with application and virtual resource metrics, gathered within the tenant domain, for identifying overloading resources in a distributed application context. The evaluation using Apache Storm as a case study, has shown that our proposal is able to identify a node experiencing multi-tenancy interference of at least 6%, with less than 1% false-positive or false-negative rates, regardless of the affected resource. Nonetheless, our model was able to generalize the multi-tenancy interference behavior based on private cloud testbed monitoring, for different hardware configurations. Thus, a system administrator can monitor an application in a public cloud provider, without possessing any hardware-level performance metrics.
{"title":"A Machine Learning Auditing Model for Detection of Multi-Tenancy Issues Within Tenant Domain","authors":"Cleverton Vicentini, A. Santin, E. Viegas, Vilmar Abreu","doi":"10.1109/CCGRID.2018.00081","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00081","url":null,"abstract":"Cloud computing is intrinsically based on multi-tenancy, which enables a physical host to be shared amongst several tenants (customers). In this context, for several reasons, a cloud provider may overload the physical machine by hosting more tenants that it can adequately handle. In such a case, a tenant may experience application performance issues. However, the tenant is not able to identify the causes, since most cloud providers do not provide performance metrics for customer monitoring, or when they do, the metrics can be biased. This study proposes a two-tier auditing model for the identification of multi-tenancy issues within the tenant domain. Our proposal relies on machine learning techniques fed with application and virtual resource metrics, gathered within the tenant domain, for identifying overloading resources in a distributed application context. The evaluation using Apache Storm as a case study, has shown that our proposal is able to identify a node experiencing multi-tenancy interference of at least 6%, with less than 1% false-positive or false-negative rates, regardless of the affected resource. Nonetheless, our model was able to generalize the multi-tenancy interference behavior based on private cloud testbed monitoring, for different hardware configurations. Thus, a system administrator can monitor an application in a public cloud provider, without possessing any hardware-level performance metrics.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"286 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124565354","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00027
Thomas B. Rolinger, T. Simon, Christopher D. Krieger
Applications for deep learning and big data analytics have compute and memory requirements that exceed the limits of a single GPU. However, effectively scaling out an application to multiple GPUs is challenging due to the complexities of communication between the GPUs, particularly for collective communication with irregular message sizes. In this work, we provide a performance evaluation of the Allgatherv routine on multi-GPU systems, focusing on GPU network topology and the communication library used. We present results from the OSU-micro benchmark as well as conduct a case study for sparse tensor factorization, one application that uses Allgatherv with highly irregular message sizes. We extend our existing tensor factorization tool to run on systems with different node counts and varying number of GPUs per node. We then evaluate the communication performance of our tool when using traditional MPI, CUDA-aware MVAPICH and NCCL across a suite of real-world data sets on three different systems: a 16-node cluster with one GPU per node, NVIDIA's DGX-1 with 8 GPUs and Cray's CS-Storm with 16 GPUs. Our results show that irregularity in the tensor data sets produce trends that contradict those in the OSU micro-benchmark, as well as trends that are absent from the benchmark.
{"title":"An Empirical Evaluation of Allgatherv on Multi-GPU Systems","authors":"Thomas B. Rolinger, T. Simon, Christopher D. Krieger","doi":"10.1109/CCGRID.2018.00027","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00027","url":null,"abstract":"Applications for deep learning and big data analytics have compute and memory requirements that exceed the limits of a single GPU. However, effectively scaling out an application to multiple GPUs is challenging due to the complexities of communication between the GPUs, particularly for collective communication with irregular message sizes. In this work, we provide a performance evaluation of the Allgatherv routine on multi-GPU systems, focusing on GPU network topology and the communication library used. We present results from the OSU-micro benchmark as well as conduct a case study for sparse tensor factorization, one application that uses Allgatherv with highly irregular message sizes. We extend our existing tensor factorization tool to run on systems with different node counts and varying number of GPUs per node. We then evaluate the communication performance of our tool when using traditional MPI, CUDA-aware MVAPICH and NCCL across a suite of real-world data sets on three different systems: a 16-node cluster with one GPU per node, NVIDIA's DGX-1 with 8 GPUs and Cray's CS-Storm with 16 GPUs. Our results show that irregularity in the tensor data sets produce trends that contradict those in the OSU micro-benchmark, as well as trends that are absent from the benchmark.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"213 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134572638","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00038
A. Segalini, Dino Lopez Pacheco, Quentin Jacquemart
In Data Centers (DCs), an abundance of virtual machines (VMs) remain idle due to network services awaiting for incoming connections, or due to established-and-idling sessions. These VMs lead to wastage of RAM – the scarcest resource in DCs – as they lock their allocated memory. In this paper, we introduce SEaMLESS, a solution designed to (i) transform fully-fledged idle VMs into lightweight and resourceless virtual network functions (VNFs), then (ii) reduces the allocated memory to those idle VMs. By replacing idle VMs with VNFs, SEaMLESS provides fast VM restoration upon user activity detection, thereby introducing limited impact on the Quality of Experience (QoE). Our results show that SEaMLESS can consolidate hundreds of VMs as VNFs onto one single machine. SEaMLESS is thus able to release the majority of the memory allocated to idle VMs. This freed memory can then be reassigned to new VMs, or lead to massive consolidation, to enable a better utilization of DC resources.
在数据中心(dc)中，由于网络服务等待传入的连接，或者由于建立和空闲会话，大量虚拟机(vm)处于空闲状态。这些vm会导致RAM(数据中心中最稀缺的资源)的浪费，因为它们会锁定已分配的内存。在本文中，我们介绍了SEaMLESS，一个旨在(i)将完全空闲的vm转换为轻量级和无资源的虚拟网络功能(VNFs)的解决方案，然后(ii)减少分配给这些空闲vm的内存。通过将空闲的虚拟机替换为VNFs, SEaMLESS可以在检测到用户活动时快速恢复虚拟机，从而减少对QoE (Quality of Experience)的影响。我们的结果表明，SEaMLESS可以将数百个vm作为VNFs整合到一台机器上。因此，SEaMLESS能够释放分配给空闲vm的大部分内存。然后可以将释放的内存重新分配给新的vm，或者进行大规模整合，以便更好地利用DC资源。
{"title":"Towards Massive Consolidation in Data Centers with SEaMLESS","authors":"A. Segalini, Dino Lopez Pacheco, Quentin Jacquemart","doi":"10.1109/CCGRID.2018.00038","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00038","url":null,"abstract":"In Data Centers (DCs), an abundance of virtual machines (VMs) remain idle due to network services awaiting for incoming connections, or due to established-and-idling sessions. These VMs lead to wastage of RAM – the scarcest resource in DCs – as they lock their allocated memory. In this paper, we introduce SEaMLESS, a solution designed to (i) transform fully-fledged idle VMs into lightweight and resourceless virtual network functions (VNFs), then (ii) reduces the allocated memory to those idle VMs. By replacing idle VMs with VNFs, SEaMLESS provides fast VM restoration upon user activity detection, thereby introducing limited impact on the Quality of Experience (QoE). Our results show that SEaMLESS can consolidate hundreds of VMs as VNFs onto one single machine. SEaMLESS is thus able to release the majority of the memory allocated to idle VMs. This freed memory can then be reassigned to new VMs, or lead to massive consolidation, to enable a better utilization of DC resources.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"117 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131989413","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00025
K. Uehara, Yu Xiang, Y. Chen, M. Hiltunen, Kaustubh R. Joshi, R. Schlichting
The explosive growth of data due to the increasing adoption of cloud technologies in the enterprise has created a strong demand for more flexible, cost-effective, and scalable storage solutions. Many storage systems, however, are not well matched to the workloads they service due to the difficulty of configuring the storage system optimally a priori with only approximate knowledge of the workload characteristics. This paper shows how cloud-based orchestration can be leveraged to create flexible storage solutions that use continuous adaptation to tailor themselves to their target application workloads, and in doing so, provide superior performance, cost, and scalability over traditional fixed designs. To demonstrate this approach, we have built "SuperCell," a Ceph-based distributed storage solution with a recommendation engine for the storage configuration. SuperCell provides storage operators with real-time recommendations on how to reconfigure the storage system to optimize its performance, cost, and efficiency based on statistical storage modeling and data analysis of the actual workload. Using real cloud storage workloads, we experimentally demonstrate that SuperCell reduces the cost of storage systems by up to 48%, while meeting service level agreement (SLA) 99% of the time, a level that any static design fails to meet for the workloads.
{"title":"SuperCell: Adaptive Software-Defined Storage for Cloud Storage Workloads","authors":"K. Uehara, Yu Xiang, Y. Chen, M. Hiltunen, Kaustubh R. Joshi, R. Schlichting","doi":"10.1109/CCGRID.2018.00025","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00025","url":null,"abstract":"The explosive growth of data due to the increasing adoption of cloud technologies in the enterprise has created a strong demand for more flexible, cost-effective, and scalable storage solutions. Many storage systems, however, are not well matched to the workloads they service due to the difficulty of configuring the storage system optimally a priori with only approximate knowledge of the workload characteristics. This paper shows how cloud-based orchestration can be leveraged to create flexible storage solutions that use continuous adaptation to tailor themselves to their target application workloads, and in doing so, provide superior performance, cost, and scalability over traditional fixed designs. To demonstrate this approach, we have built \"SuperCell,\" a Ceph-based distributed storage solution with a recommendation engine for the storage configuration. SuperCell provides storage operators with real-time recommendations on how to reconfigure the storage system to optimize its performance, cost, and efficiency based on statistical storage modeling and data analysis of the actual workload. Using real cloud storage workloads, we experimentally demonstrate that SuperCell reduces the cost of storage systems by up to 48%, while meeting service level agreement (SLA) 99% of the time, a level that any static design fails to meet for the workloads.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"72 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131582213","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00065
Yi Zhou, Shubbhi Taneja, Mohammed I. Alghamdi, X. Qin
The goal of this study is to optimize energy efficiency of database clusters through prefetching and caching strategies. We design a workload-skewness scheme to collectively manage a set of hot and cold nodes in a database cluster system. The prefetching mechanism fetches popular data tables to the hot nodes while keeping unpopular data in cold nodes. We leverage a power management module to aggressively turn cold nodes in the low-power mode to conserve energy consumption. We construct a prefetching model and an energy-saving model to govern the power management module in database lusters. The energy-efficient prefetching and caching mechanism is conducive to cutting back the number of power-state transitions, thereby offering high energy efficiency. We systematically evaluate energy conservation technique in the process of managing, fetching, and storing data on clusters supporting database applications. Our experimental results show that our prefetching/caching solution significantly improves energy efficiency of the existing PostgreSQL system.
{"title":"Improving Energy Efficiency of Database Clusters Through Prefetching and Caching","authors":"Yi Zhou, Shubbhi Taneja, Mohammed I. Alghamdi, X. Qin","doi":"10.1109/CCGRID.2018.00065","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00065","url":null,"abstract":"The goal of this study is to optimize energy efficiency of database clusters through prefetching and caching strategies. We design a workload-skewness scheme to collectively manage a set of hot and cold nodes in a database cluster system. The prefetching mechanism fetches popular data tables to the hot nodes while keeping unpopular data in cold nodes. We leverage a power management module to aggressively turn cold nodes in the low-power mode to conserve energy consumption. We construct a prefetching model and an energy-saving model to govern the power management module in database lusters. The energy-efficient prefetching and caching mechanism is conducive to cutting back the number of power-state transitions, thereby offering high energy efficiency. We systematically evaluate energy conservation technique in the process of managing, fetching, and storing data on clusters supporting database applications. Our experimental results show that our prefetching/caching solution significantly improves energy efficiency of the existing PostgreSQL system.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"26 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114874299","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00071
Vito Giovanni Castellana, Marco Minutoli
The unprecedented amount of data that needs to be processed in emerging data analytics applications poses novel challenges to industry and academia. Scalability and high performance become more than a desirable feature because, due to the scale and the nature of the problems, they draw the line between what is achievable and what is unfeasible. In this paper, we propose SHAD, the Scalable High-performance Algorithms and Data-structures library. SHAD adopts a modular design that confines low level details and promotes reuse. SHAD's core is built on an Abstract Runtime Interface which enhances portability and identifies the minimal set of features of the underlying system required by the framework. The core library includes common data-structures such as: Array, Vector, Map and Set. These are designed to accommodate significant amount of data which can be accessed in massively parallel environments, and used as building blocks for SHAD extensions, i.e. higher level software libraries. We have validated and evaluated our design with a performance and scalability study of the core components of the library. We have validated the design flexibility by proposing a Graph Library as an example of SHAD extension, which implements two different graph data-structures; we evaluate their performance with a set of graph applications. Experimental results show that the approach is promising in terms of both performance and scalability. On a distributed system with 320 cores, SHAD Arrays are able to sustain a throughput of 65 billion operations per second, while SHAD Maps sustain 1 billion of operations per second. Algorithms implemented using the Graph Library exhibit performance and scalability comparable to a custom solution, but with smaller development effort.
{"title":"SHAD: The Scalable High-Performance Algorithms and Data-Structures Library","authors":"Vito Giovanni Castellana, Marco Minutoli","doi":"10.1109/CCGRID.2018.00071","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00071","url":null,"abstract":"The unprecedented amount of data that needs to be processed in emerging data analytics applications poses novel challenges to industry and academia. Scalability and high performance become more than a desirable feature because, due to the scale and the nature of the problems, they draw the line between what is achievable and what is unfeasible. In this paper, we propose SHAD, the Scalable High-performance Algorithms and Data-structures library. SHAD adopts a modular design that confines low level details and promotes reuse. SHAD's core is built on an Abstract Runtime Interface which enhances portability and identifies the minimal set of features of the underlying system required by the framework. The core library includes common data-structures such as: Array, Vector, Map and Set. These are designed to accommodate significant amount of data which can be accessed in massively parallel environments, and used as building blocks for SHAD extensions, i.e. higher level software libraries. We have validated and evaluated our design with a performance and scalability study of the core components of the library. We have validated the design flexibility by proposing a Graph Library as an example of SHAD extension, which implements two different graph data-structures; we evaluate their performance with a set of graph applications. Experimental results show that the approach is promising in terms of both performance and scalability. On a distributed system with 320 cores, SHAD Arrays are able to sustain a throughput of 65 billion operations per second, while SHAD Maps sustain 1 billion of operations per second. Algorithms implemented using the Graph Library exhibit performance and scalability comparable to a custom solution, but with smaller development effort.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"38 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124036661","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00056
Dung Nguyen, André Luckow, Edward B. Duffy, Ken E. Kennedy, A. Apon
This paper presents a systematic evaluation of Amazon Kinesis and Apache Kafka for meeting highly demanding application requirements. Results show that Kinesis and Kafka can provide high reliability, performance and scalability. Cost and performance trade-offs of Kinesis and Kafka are presented for a variety of application data rates, resource utilization, and resource configurations.
{"title":"Evaluation of Highly Available Cloud Streaming Systems for Performance and Price","authors":"Dung Nguyen, André Luckow, Edward B. Duffy, Ken E. Kennedy, A. Apon","doi":"10.1109/CCGRID.2018.00056","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00056","url":null,"abstract":"This paper presents a systematic evaluation of Amazon Kinesis and Apache Kafka for meeting highly demanding application requirements. Results show that Kinesis and Kafka can provide high reliability, performance and scalability. Cost and performance trade-offs of Kinesis and Kafka are presented for a variety of application data rates, resource utilization, and resource configurations.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"15 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121050171","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00043
Li-Yung Ho, Jan-Jan Wu, Pangfeng Liu
Deep learning is now the most promising approach to develop human-intelligent computer systems. To speedup the development of neural networks, researchers have designed many distributed learning algorithms to facilitate the training process. In these algorithms, people use a constant to indicate the communication period for model/gradient exchange. We find that this type of communication pattern could incur unnecessary and inefficient data transmission for some training methods e.g., elastic SGD and gossiping SGD. In this paper, we propose an adaptive communication method to improve the performance of gossiping SGD. Instead of using a fixed period for model exchange, we exchange the models with other machines according to the change of the local model. This makes the communication more efficient and thus improves the performance. The experiment results show that our method reduces the communication traffic by 92%, which results in 52% reduction in training time while preserving the prediction accuracy compared with gossiping SGD.
{"title":"Adaptive Communication for Distributed Deep Learning on Commodity GPU Cluster","authors":"Li-Yung Ho, Jan-Jan Wu, Pangfeng Liu","doi":"10.1109/CCGRID.2018.00043","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00043","url":null,"abstract":"Deep learning is now the most promising approach to develop human-intelligent computer systems. To speedup the development of neural networks, researchers have designed many distributed learning algorithms to facilitate the training process. In these algorithms, people use a constant to indicate the communication period for model/gradient exchange. We find that this type of communication pattern could incur unnecessary and inefficient data transmission for some training methods e.g., elastic SGD and gossiping SGD. In this paper, we propose an adaptive communication method to improve the performance of gossiping SGD. Instead of using a fixed period for model exchange, we exchange the models with other machines according to the change of the local model. This makes the communication more efficient and thus improves the performance. The experiment results show that our method reduces the communication traffic by 92%, which results in 52% reduction in training time while preserving the prediction accuracy compared with gossiping SGD.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"20 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121177898","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00028
Aleksandra Kuzmanovska, H. V. D. Bogert, R. H. Mak, D. Epema
When multiple data-processing frameworks with time-varying workloads are simultaneously present in a single cluster or data-center, an apparent goal is to have them experience equal performance, expressed in whatever performance metrics are applicable. In modern data-center environments, Two-Level Schedulers (TLSs) that leave the scheduling of individual jobs to the schedulers within the data-processing frameworks are typically used for managing the resources of data-processing frameworks. Two such TLSs with opposite designs are Mesos and Koala-F. Mesos employs fine-grained resource allocation and aims at Dominant Resource Fairness (DRF) among framework instances by offering resources to them for the duration of a single task. In contrast, Koala-F aims at performance fairness among framework instances by employing dynamic coarse-grained resource allocation of sets of complete nodes based on performance feedback from individual instances. The goal of this paper is to explore the trade-offs between these two TLS designs when trying to achieve performance balance among frameworks. We select Apache Spark as a representative of data-processing frameworks, and perform experiments on a modest-sized cluster, using jobs chosen from commonly used data-processing benchmarks. Our results reveal that achieving performance balance among framework instances is a challenge for both TLS designs, despite their opposite design choices. Moreover, we exhibit design flaws in the DRF allocation policy that prevent Mesos from achieving performance balance. Finally, to remedy these flaws, we propose a feedback controller for Mesos that dynamically adapts framework weights, as used in Weighted DRF (W-DRF), based on their performance.
{"title":"Achieving Performance Balance Among Spark Frameworks with Two-Level Schedulers","authors":"Aleksandra Kuzmanovska, H. V. D. Bogert, R. H. Mak, D. Epema","doi":"10.1109/CCGRID.2018.00028","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00028","url":null,"abstract":"When multiple data-processing frameworks with time-varying workloads are simultaneously present in a single cluster or data-center, an apparent goal is to have them experience equal performance, expressed in whatever performance metrics are applicable. In modern data-center environments, Two-Level Schedulers (TLSs) that leave the scheduling of individual jobs to the schedulers within the data-processing frameworks are typically used for managing the resources of data-processing frameworks. Two such TLSs with opposite designs are Mesos and Koala-F. Mesos employs fine-grained resource allocation and aims at Dominant Resource Fairness (DRF) among framework instances by offering resources to them for the duration of a single task. In contrast, Koala-F aims at performance fairness among framework instances by employing dynamic coarse-grained resource allocation of sets of complete nodes based on performance feedback from individual instances. The goal of this paper is to explore the trade-offs between these two TLS designs when trying to achieve performance balance among frameworks. We select Apache Spark as a representative of data-processing frameworks, and perform experiments on a modest-sized cluster, using jobs chosen from commonly used data-processing benchmarks. Our results reveal that achieving performance balance among framework instances is a challenge for both TLS designs, despite their opposite design choices. Moreover, we exhibit design flaws in the DRF allocation policy that prevent Mesos from achieving performance balance. Finally, to remedy these flaws, we propose a feedback controller for Mesos that dynamically adapts framework weights, as used in Weighted DRF (W-DRF), based on their performance.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116725946","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}
Pub Date : 2018-05-01DOI: 10.1109/CCGRID.2018.00007
G. Laccetti, M. Lapegna, R. Montella
Concurrent data structures are widely used in many software stack levels, ranging from high level parallel scientific applications to low level operating systems. The key issue of these objects is their concurrent use by several computing units (threads or process) so that the design of these structures is much more difficult compared to their sequential counterpart, because of their extremely dynamic nature requiring protocols to ensure data consistency, with a significant cost overhead. At this regard, several studies emphasize a tension between the needs of sequential correctness of the concurrent data structures and scalability of the algorithms, and in many cases it is evident the need to rethink the data structure design, using approaches based on randomization and/or redistribution techniques in order to fully exploit the computational power of the recent computing environments. The problem is grown in importance with the new generation High Performance Computing systems aimed to achieve extreme performance. It is easy to observe that such systems are based on heterogeneous architectures integrating several independent nodes in the form of clusters or MPP systems, where each node is composed by powerful computing elements (CPU core, GPUs or other acceleration devices) sharing resources in a single node. These systems therefore make massive use of communication libraries to exchange data among the nodes, as well as other tools for the management of the shared resources inside a single node. For such a reason, the development of algorithms and scientific software for dynamic data structures on these heterogeneous systems implies a suitable combination of several methodologies and tools to deal with the different kinds of parallelism corresponding to each specific device, so that to be aware of the underlying platform. The present work is aimed to introduce a scalable model to manage a special class of dynamic data structure known as heap based priority queue (or simply heap) on these heterogeneous architectures. A heap is generally used when the applications needs set of data not requiring a complete ordering, but only the access to some items tagged with high priority. In order to ensure a tradeoff between the correct access to high priority items by the several computing units with a low communication and synchronization overhead, a suitable reorganization of the heap is needed. More precisely we introduce a unified scalable model that can be used, with no modifications, to redeploy the items of a heap both in message passing environments (such as clusters and or MMP multicomputers with several nodes) as well as in shared memory environments (such as CPUs and multiprocessors with several cores) with an overhead independent of the number of computing units. Computational results related to the application of the proposed strategy on some numerical case studies are presented for different types of computing environments.
{"title":"A Scalable Unified Model for Dynamic Data Structures in Message Passing (Clusters) and Shared Memory (multicore CPUs) Computing environments","authors":"G. Laccetti, M. Lapegna, R. Montella","doi":"10.1109/CCGRID.2018.00007","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00007","url":null,"abstract":"Concurrent data structures are widely used in many software stack levels, ranging from high level parallel scientific applications to low level operating systems. The key issue of these objects is their concurrent use by several computing units (threads or process) so that the design of these structures is much more difficult compared to their sequential counterpart, because of their extremely dynamic nature requiring protocols to ensure data consistency, with a significant cost overhead. At this regard, several studies emphasize a tension between the needs of sequential correctness of the concurrent data structures and scalability of the algorithms, and in many cases it is evident the need to rethink the data structure design, using approaches based on randomization and/or redistribution techniques in order to fully exploit the computational power of the recent computing environments. The problem is grown in importance with the new generation High Performance Computing systems aimed to achieve extreme performance. It is easy to observe that such systems are based on heterogeneous architectures integrating several independent nodes in the form of clusters or MPP systems, where each node is composed by powerful computing elements (CPU core, GPUs or other acceleration devices) sharing resources in a single node. These systems therefore make massive use of communication libraries to exchange data among the nodes, as well as other tools for the management of the shared resources inside a single node. For such a reason, the development of algorithms and scientific software for dynamic data structures on these heterogeneous systems implies a suitable combination of several methodologies and tools to deal with the different kinds of parallelism corresponding to each specific device, so that to be aware of the underlying platform. The present work is aimed to introduce a scalable model to manage a special class of dynamic data structure known as heap based priority queue (or simply heap) on these heterogeneous architectures. A heap is generally used when the applications needs set of data not requiring a complete ordering, but only the access to some items tagged with high priority. In order to ensure a tradeoff between the correct access to high priority items by the several computing units with a low communication and synchronization overhead, a suitable reorganization of the heap is needed. More precisely we introduce a unified scalable model that can be used, with no modifications, to redeploy the items of a heap both in message passing environments (such as clusters and or MMP multicomputers with several nodes) as well as in shared memory environments (such as CPUs and multiprocessors with several cores) with an overhead independent of the number of computing units. Computational results related to the application of the proposed strategy on some numerical case studies are presented for different types of computing environments.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"378 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126972784","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}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：