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}
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.2018.00029
Subarna Chatterjee, C. Morin
With the advent of the Internet of Things (IoT), data stream processing have gained increased attention due to the ever-increasing need to process heterogeneous and voluminous data streams. This work addresses the problem of selecting a correct stream processing framework for a given application to be executed within a specific physical infrastructure. For this purpose, we focus on a thorough comparative analysis of three data stream processing platforms – Apache Flink, Apache Storm, and Twitter Heron (the enhanced version of Apache Storm), that are chosen based on their potential to process both streams and batches in real-time. The goal of the work is to enlighten the cloud-clients and the cloud-providers with the knowledge of the choice of the resource-efficient and requirement-adaptive streaming platform for a given application so that they can plan during allocation or assignment of Virtual Machines for application execution. For the comparative performance analysis of the chosen platforms, we have experimented using 8-node clusters on Grid5000 experimentation testbed and have selected a wide variety of applications ranging from a conventional benchmark to sensor-based IoT application and statistical batch processing application. In addition to the various performance metrics related to the elasticity and resource usage of the platforms, this work presents a comparative study of the “green-ness” of the streaming platforms by analyzing their power consumption – one of the first attempts of its kind. The obtained results are thoroughly analyzed to illustrate the functional behavior of these platforms under different computing scenarios.
{"title":"Experimental Study on the Performance and Resource Utilization of Data Streaming Frameworks","authors":"Subarna Chatterjee, C. Morin","doi":"10.1109/CCGRID.2018.00029","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00029","url":null,"abstract":"With the advent of the Internet of Things (IoT), data stream processing have gained increased attention due to the ever-increasing need to process heterogeneous and voluminous data streams. This work addresses the problem of selecting a correct stream processing framework for a given application to be executed within a specific physical infrastructure. For this purpose, we focus on a thorough comparative analysis of three data stream processing platforms – Apache Flink, Apache Storm, and Twitter Heron (the enhanced version of Apache Storm), that are chosen based on their potential to process both streams and batches in real-time. The goal of the work is to enlighten the cloud-clients and the cloud-providers with the knowledge of the choice of the resource-efficient and requirement-adaptive streaming platform for a given application so that they can plan during allocation or assignment of Virtual Machines for application execution. For the comparative performance analysis of the chosen platforms, we have experimented using 8-node clusters on Grid5000 experimentation testbed and have selected a wide variety of applications ranging from a conventional benchmark to sensor-based IoT application and statistical batch processing application. In addition to the various performance metrics related to the elasticity and resource usage of the platforms, this work presents a comparative study of the “green-ness” of the streaming platforms by analyzing their power consumption – one of the first attempts of its kind. The obtained results are thoroughly analyzed to illustrate the functional behavior of these platforms under different computing scenarios.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"111 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":"123834208","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.2018.00073
G. Vernik, M. Factor, E. K. Kolodner, P. Michiardi, Effi Ofer, Francesco Pace
Until now object storage has not been a first-class citizen of the Apache Hadoop ecosystem including Apache Spark. Hadoop connectors to object storage have been based on file semantics, an impedance mismatch, which leads to low performance and the need for an additional consistent storage system to achieve fault tolerance. In particular, Hadoop depends on its underlying storage system and its associated connector for fault tolerance and allowing speculative execution. However, these characteristics are obtained through file operations that are not native for object storage, and are both costly and not atomic. As a result these connectors are not efficient and more importantly they cannot help with fault tolerance for object storage. We introduce Stocator, whose novel algorithm achieves both high performance and fault tolerance by taking advantage of object storage semantics. This greatly decreases the number of operations on object storage as well as enabling a much simpler approach to dealing with the eventually consistent semantics typical of object storage. We have implemented Stocator and shared it in open source. Performance testing with Apache Spark shows that it can be 18 times faster for write intensive workloads and can perform 30 times fewer operations on object storage than the legacy Hadoop connectors, reducing costs both for the client and the object storage service provider.
{"title":"Stocator: Providing High Performance and Fault Tolerance for Apache Spark Over Object Storage","authors":"G. Vernik, M. Factor, E. K. Kolodner, P. Michiardi, Effi Ofer, Francesco Pace","doi":"10.1109/CCGRID.2018.00073","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00073","url":null,"abstract":"Until now object storage has not been a first-class citizen of the Apache Hadoop ecosystem including Apache Spark. Hadoop connectors to object storage have been based on file semantics, an impedance mismatch, which leads to low performance and the need for an additional consistent storage system to achieve fault tolerance. In particular, Hadoop depends on its underlying storage system and its associated connector for fault tolerance and allowing speculative execution. However, these characteristics are obtained through file operations that are not native for object storage, and are both costly and not atomic. As a result these connectors are not efficient and more importantly they cannot help with fault tolerance for object storage. We introduce Stocator, whose novel algorithm achieves both high performance and fault tolerance by taking advantage of object storage semantics. This greatly decreases the number of operations on object storage as well as enabling a much simpler approach to dealing with the eventually consistent semantics typical of object storage. We have implemented Stocator and shared it in open source. Performance testing with Apache Spark shows that it can be 18 times faster for write intensive workloads and can perform 30 times fewer operations on object storage than the legacy Hadoop connectors, reducing costs both for the client and the object storage service provider.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"197 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":"121107609","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.2018.00080
Daniel Mawhirter, Bo Wu, D. Mehta, Chao Ai
Graphlet counting is a methodology for detecting local structural properties of large graphs that has been in use for over a decade. Despite tremendous effort in optimizing its performance, even 3- and 4-node graphlet counting routines may run for hours or days on highly optimized systems. In this paper, we describe how a synergistic combination of approximate computing with parallel computing can result in multiplicative performance improvements in graphlet counting runtimes with minimal and controllable loss of accuracy. Specifically, we describe two novel techniques, multi-phased sampling for statistical accuracy guarantees and cost-aware sampling to further improve performance on multi-machine runs, which reduce the query time on large graphs from tens of hours to several minutes or seconds with only <1% relative error.
{"title":"ApproxG: Fast Approximate Parallel Graphlet Counting Through Accuracy Control","authors":"Daniel Mawhirter, Bo Wu, D. Mehta, Chao Ai","doi":"10.1109/CCGRID.2018.00080","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00080","url":null,"abstract":"Graphlet counting is a methodology for detecting local structural properties of large graphs that has been in use for over a decade. Despite tremendous effort in optimizing its performance, even 3- and 4-node graphlet counting routines may run for hours or days on highly optimized systems. In this paper, we describe how a synergistic combination of approximate computing with parallel computing can result in multiplicative performance improvements in graphlet counting runtimes with minimal and controllable loss of accuracy. Specifically, we describe two novel techniques, multi-phased sampling for statistical accuracy guarantees and cost-aware sampling to further improve performance on multi-machine runs, which reduce the query time on large graphs from tens of hours to several minutes or seconds with only <1% relative error.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"325 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":"123213988","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.2018.00020
Pedro Raminhas, S. Issa, P. Romano
Transactional Memory (TM) is an emerging paradigm that promises to significantly ease the development of parallel programs. Hybrid TM (HyTM) is probably the most promising implementation of the TM abstraction, which seeks to combine the high efficiency of hardware implementations (HTM) with the robustness and flexibility of software-based ones (STM). Unfortunately, though, existing Hybrid TM systems are known to suffer from high overheads to guarantee correct synchronization between concurrent transactions executing in hardware and software. This article introduces DMP-TM (Dynamic Memory Partitioning-TM), a novel HyTM algorithm that exploits, to the best of our knowledge for the first time in the literature, the idea of leveraging operating system-level memory protection mechanisms to detect conflicts between HTM and STM transactions. This innovative design allows for employing highly scalable STM implementations, while avoiding instrumentation on the HTM path. This allows DMP-TM to achieve up to ~ 20× speedups compared to state of the art Hybrid TM solutions in uncontended workloads. Further, thanks to the use of simple and lightweight self-tuning mechanisms, DMP-TM achieves robust performance even in unfavourable workload that exhibits high contention between the STM and HTM path.
{"title":"Enhancing Efficiency of Hybrid Transactional Memory Via Dynamic Data Partitioning Schemes","authors":"Pedro Raminhas, S. Issa, P. Romano","doi":"10.1109/CCGRID.2018.00020","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00020","url":null,"abstract":"Transactional Memory (TM) is an emerging paradigm that promises to significantly ease the development of parallel programs. Hybrid TM (HyTM) is probably the most promising implementation of the TM abstraction, which seeks to combine the high efficiency of hardware implementations (HTM) with the robustness and flexibility of software-based ones (STM). Unfortunately, though, existing Hybrid TM systems are known to suffer from high overheads to guarantee correct synchronization between concurrent transactions executing in hardware and software. This article introduces DMP-TM (Dynamic Memory Partitioning-TM), a novel HyTM algorithm that exploits, to the best of our knowledge for the first time in the literature, the idea of leveraging operating system-level memory protection mechanisms to detect conflicts between HTM and STM transactions. This innovative design allows for employing highly scalable STM implementations, while avoiding instrumentation on the HTM path. This allows DMP-TM to achieve up to ~ 20× speedups compared to state of the art Hybrid TM solutions in uncontended workloads. Further, thanks to the use of simple and lightweight self-tuning mechanisms, DMP-TM achieves robust performance even in unfavourable workload that exhibits high contention between the STM and HTM path.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"1 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":"128486893","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.2018.00070
Wenqi Liu, Hongxiang Li, Bin Xie
Recently, large-scale networks attract significant attention to analyze and extract the hidden information of big data. Toward this end, graph embedding is a method to embed a high dimensional graph into a much lower dimensional vector space while maximally preserving the structural information of the original network. However, effective graph embedding is particularly challenging when massive graph data are generated and processed for real-time applications. In this paper, we address this challenge and propose a new real-time and distributed graph embedding algorithm (RTDGE) that is capable of distributively embedding a large-scale graph in a streaming fashion. Specifically, our RTDGE consists of the following components: (1) a graph partition scheme that divides all edges into distinct subgraphs, where vertices are associated with edges and may belong to several subgraphs; (2) a dynamic negative sampling (DNS) method that updates the embedded vectors in real-time; and (3) an unsupervised global aggregation scheme that combines all locally embedded vectors into a global vector space. Furthermore, we also build a real-time distributed graph embedding platform based on Apache Kafka and Apache Storm. Extensive experimental results show that RTDGE outperforms existing solutions in terms of graph embedding efficiency and accuracy.
{"title":"Real-Time Graph Partition and Embedding of Large Network","authors":"Wenqi Liu, Hongxiang Li, Bin Xie","doi":"10.1109/CCGRID.2018.00070","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00070","url":null,"abstract":"Recently, large-scale networks attract significant attention to analyze and extract the hidden information of big data. Toward this end, graph embedding is a method to embed a high dimensional graph into a much lower dimensional vector space while maximally preserving the structural information of the original network. However, effective graph embedding is particularly challenging when massive graph data are generated and processed for real-time applications. In this paper, we address this challenge and propose a new real-time and distributed graph embedding algorithm (RTDGE) that is capable of distributively embedding a large-scale graph in a streaming fashion. Specifically, our RTDGE consists of the following components: (1) a graph partition scheme that divides all edges into distinct subgraphs, where vertices are associated with edges and may belong to several subgraphs; (2) a dynamic negative sampling (DNS) method that updates the embedded vectors in real-time; and (3) an unsupervised global aggregation scheme that combines all locally embedded vectors into a global vector space. Furthermore, we also build a real-time distributed graph embedding platform based on Apache Kafka and Apache Storm. Extensive experimental results show that RTDGE outperforms existing solutions in terms of graph embedding efficiency and accuracy.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"33 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":"125326117","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.2018.00061
R. Luley, Qinru Qiu
Optimizing resource utilization is a critical issue in cloud and cluster-based computing systems. In such systems, computing resources often consist of one or more GPU devices, and much research has already been conducted on means for maximizing compute resources through shared execution strategies. However, one of the most severe resource constraints in these scenarios is the data transfer channel between the host (i.e., CPU) and the device (i.e., GPU). Data transfer contention has been shown to have a significant impact on performance, yet methods for optimizing such contention have not been thoroughly studied. Techniques that have been examined make certain assumptions which limit effectiveness in the general case. In this paper, we introduce a heuristic which selectively aggregates transfers in order to maximize system performance by optimizing the transfer channel bandwidth. We compare this heuristic to traditional first-come-first-served approach, and apply Monte Carlo reinforcement learning to find an optimal policy for message aggregation. Finally, we evaluate the performance of Monte Carlo reinforcement learning with an arbitrarily-initialized policy. We demonstrate its effectiveness in learning optimal data transfer policy without detailed system characterization, which will enable a general adaptable solution for resource management of future systems.
{"title":"Optimizing Data Transfers for Improved Performance on Shared GPUs Using Reinforcement Learning","authors":"R. Luley, Qinru Qiu","doi":"10.1109/CCGRID.2018.00061","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00061","url":null,"abstract":"Optimizing resource utilization is a critical issue in cloud and cluster-based computing systems. In such systems, computing resources often consist of one or more GPU devices, and much research has already been conducted on means for maximizing compute resources through shared execution strategies. However, one of the most severe resource constraints in these scenarios is the data transfer channel between the host (i.e., CPU) and the device (i.e., GPU). Data transfer contention has been shown to have a significant impact on performance, yet methods for optimizing such contention have not been thoroughly studied. Techniques that have been examined make certain assumptions which limit effectiveness in the general case. In this paper, we introduce a heuristic which selectively aggregates transfers in order to maximize system performance by optimizing the transfer channel bandwidth. We compare this heuristic to traditional first-come-first-served approach, and apply Monte Carlo reinforcement learning to find an optimal policy for message aggregation. Finally, we evaluate the performance of Monte Carlo reinforcement learning with an arbitrarily-initialized policy. We demonstrate its effectiveness in learning optimal data transfer policy without detailed system characterization, which will enable a general adaptable solution for resource management of future systems.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"137 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":"116722027","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.2018.00049
P. Minet, É. Renault, I. Khoufi, S. Boumerdassi
Data collected from an operational Google data center during 29 days represent a very rich and very useful source of information for understanding the main features of a data center. In this paper, we highlight the strong heterogeneity of jobs. The distribution of job execution duration shows a high disparity, as well as the job waiting time before being scheduled. The resource requests in terms of CPU and memory are also analyzed. The knowledge of all these features is needed to design models of jobs, machines and resource requests that are representative of a real data center.
{"title":"Data Analysis of a Google Data Center","authors":"P. Minet, É. Renault, I. Khoufi, S. Boumerdassi","doi":"10.1109/CCGRID.2018.00049","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00049","url":null,"abstract":"Data collected from an operational Google data center during 29 days represent a very rich and very useful source of information for understanding the main features of a data center. In this paper, we highlight the strong heterogeneity of jobs. The distribution of job execution duration shows a high disparity, as well as the job waiting time before being scheduled. The resource requests in terms of CPU and memory are also analyzed. The knowledge of all these features is needed to design models of jobs, machines and resource requests that are representative of a real data center.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"14 2 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":"116822815","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 : 2018-04-28DOI: 10.1109/CCGRID.2018.00044
Sudheer Chunduri, Meysam Ghaffari, M. S. Lahijani, A. Srinivasan, S. Namilae
Numerical simulations are used to analyze the effectiveness of alternate public policy choices in limiting the spread of infections. In practice, it is usually not feasible to predict their precise impacts due to inherent uncertainties, especially at the early stages of an epidemic. One option is to parameterize the sources of uncertainty and carry out a parameter sweep to identify their robustness under a variety of possible scenarios. The Self Propelled Entity Dynamics (SPED) model has used this approach successfully to analyze the robustness of different airline boarding and deplaning procedures. However, the time taken by this approach is too large to answer questions raised during the course of a decision meeting. In this paper, we use a modified approach that pre-computes simulations of passenger movement, performing only the disease-specific analysis in real time. A novel contribution of this paper lies in using a low discrepancy sequence (LDS) in the parameter sweep, and demonstrating that it can lead to a reduction in analysis time by one to three orders of magnitude over the conventional lattice-based parameter sweep. However, its parallelization suffers from greater load imbalance than the conventional approach. We examine this and relate it to number-theoretic properties of the LDS. We then propose solutions to this problem. Our approach and analysis are applicable to other parameter sweep problems too. The primary contributions of this paper lie in the new approach of low discrepancy parameter sweep and in exploring solutions to challenges in its parallelization, evaluated in the context of an important public health application.
{"title":"Parallel Low Discrepancy Parameter Sweep for Public Health Policy","authors":"Sudheer Chunduri, Meysam Ghaffari, M. S. Lahijani, A. Srinivasan, S. Namilae","doi":"10.1109/CCGRID.2018.00044","DOIUrl":"https://doi.org/10.1109/CCGRID.2018.00044","url":null,"abstract":"Numerical simulations are used to analyze the effectiveness of alternate public policy choices in limiting the spread of infections. In practice, it is usually not feasible to predict their precise impacts due to inherent uncertainties, especially at the early stages of an epidemic. One option is to parameterize the sources of uncertainty and carry out a parameter sweep to identify their robustness under a variety of possible scenarios. The Self Propelled Entity Dynamics (SPED) model has used this approach successfully to analyze the robustness of different airline boarding and deplaning procedures. However, the time taken by this approach is too large to answer questions raised during the course of a decision meeting. In this paper, we use a modified approach that pre-computes simulations of passenger movement, performing only the disease-specific analysis in real time. A novel contribution of this paper lies in using a low discrepancy sequence (LDS) in the parameter sweep, and demonstrating that it can lead to a reduction in analysis time by one to three orders of magnitude over the conventional lattice-based parameter sweep. However, its parallelization suffers from greater load imbalance than the conventional approach. We examine this and relate it to number-theoretic properties of the LDS. We then propose solutions to this problem. Our approach and analysis are applicable to other parameter sweep problems too. The primary contributions of this paper lie in the new approach of low discrepancy parameter sweep and in exploring solutions to challenges in its parallelization, evaluated in the context of an important public health application.","PeriodicalId":321027,"journal":{"name":"2018 18th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID)","volume":"53 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133561074","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：