Cy P. Chan, J. Bachan, J. Kenny, Jeremiah J. Wilke, V. Beckner, A. Almgren, J. Bell
We introduce a topology-aware performance optimization and modeling workflow for AMR simulation that includes two new modeling tools, ProgrAMR and Mota Mapper, which interface with the BoxLib AMR framework and the SSTmacro network simulator. ProgrAMR allows us to generate and model the execution of task dependency graphs from high-level specifications of AMR-based applications, which we demonstrate by analyzing two example AMR-based multigrid solvers with varying degrees of asynchrony. Mota Mapper generates multiobjective, network topology-aware box mappings, which we apply to optimize the data layout for the example multigrid solvers. While the sensitivity of these solvers to layout and execution strategy appears to be modest for balanced scenarios, the impact of better mapping algorithms can be significant when performance is highly constrained by network hop latency. Furthermore, we show that network latency in the multigrid bottom solve is the main contributing factor preventing good scaling on exascale-class machines.
{"title":"Topology-Aware Performance Optimization and Modeling of Adaptive Mesh Refinement Codes for Exascale","authors":"Cy P. Chan, J. Bachan, J. Kenny, Jeremiah J. Wilke, V. Beckner, A. Almgren, J. Bell","doi":"10.1109/COM-HPC.2016.8","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.8","url":null,"abstract":"We introduce a topology-aware performance optimization and modeling workflow for AMR simulation that includes two new modeling tools, ProgrAMR and Mota Mapper, which interface with the BoxLib AMR framework and the SSTmacro network simulator. ProgrAMR allows us to generate and model the execution of task dependency graphs from high-level specifications of AMR-based applications, which we demonstrate by analyzing two example AMR-based multigrid solvers with varying degrees of asynchrony. Mota Mapper generates multiobjective, network topology-aware box mappings, which we apply to optimize the data layout for the example multigrid solvers. While the sensitivity of these solvers to layout and execution strategy appears to be modest for balanced scenarios, the impact of better mapping algorithms can be significant when performance is highly constrained by network hop latency. Furthermore, we show that network latency in the multigrid bottom solve is the main contributing factor preventing good scaling on exascale-class machines.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129090679","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}
R. Graham, Devendar Bureddy, Pak Lui, H. Rosenstock, G. Shainer, Gil Bloch, Dror Goldenberg, Mike Dubman, Sasha Kotchubievsky, Vladimir Koushnir, Lion Levi, Alexander Margolin, Tamir Ronen, Alexander Shpiner, O. Wertheim, E. Zahavi
Increased system size and a greater reliance on utilizing system parallelism to achieve computational needs, requires innovative system architectures to meet the simulation challenges. As a step towards a new network class of co-processors — intelligent network devices, which manipulate data traversing the data-center network, this paper describes the SHArP technology designed to offload collective operation processing to the network. This is implemented in Mellanox's SwitchIB-2 ASIC, using innetwork trees to reduce data from a group of sources, and to distribute the result. Multiple parallel jobs with several partially overlapping groups are supported each with several reduction operations in-flight. Large performance enhancements are obtained, with an improvement of a factor of 2.1 for an eight byte MPI_Allreduce() operation on 128 hosts, going from 6.01 to 2.83 microseconds. Pipelining is used for an improvement of a factor of 3.24 in the latency of a 4096 byte MPI_Allreduce() operations, declining from 46.93 to 14.48 microseconds.
{"title":"Scalable Hierarchical Aggregation Protocol (SHArP): A Hardware Architecture for Efficient Data Reduction","authors":"R. Graham, Devendar Bureddy, Pak Lui, H. Rosenstock, G. Shainer, Gil Bloch, Dror Goldenberg, Mike Dubman, Sasha Kotchubievsky, Vladimir Koushnir, Lion Levi, Alexander Margolin, Tamir Ronen, Alexander Shpiner, O. Wertheim, E. Zahavi","doi":"10.1109/COM-HPC.2016.6","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.6","url":null,"abstract":"Increased system size and a greater reliance on utilizing system parallelism to achieve computational needs, requires innovative system architectures to meet the simulation challenges. As a step towards a new network class of co-processors — intelligent network devices, which manipulate data traversing the data-center network, this paper describes the SHArP technology designed to offload collective operation processing to the network. This is implemented in Mellanox's SwitchIB-2 ASIC, using innetwork trees to reduce data from a group of sources, and to distribute the result. Multiple parallel jobs with several partially overlapping groups are supported each with several reduction operations in-flight. Large performance enhancements are obtained, with an improvement of a factor of 2.1 for an eight byte MPI_Allreduce() operation on 128 hosts, going from 6.01 to 2.83 microseconds. Pipelining is used for an improvement of a factor of 3.24 in the latency of a 4096 byte MPI_Allreduce() operations, declining from 46.93 to 14.48 microseconds.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"120823754","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}
Grey Ballard, J. Demmel, A. Gearhart, Benjamin Lipshitz, Yishai Oltchik, O. Schwartz, Sivan Toledo
Network topologies can have significant effect on the execution costs of parallel algorithms due to inter-processor communication. For particular combinations of computations and network topologies, costly network contention may inevitably become a bottleneck, even if algorithms are optimally designed so that each processor communicates as little as possible. We obtain novel contention lower bounds that are functions of the network and the computation graph parameters. For several combinations of fundamental computations and common network topologies, our new analysis improves upon previous per-processor lower bounds which only specify the number of words communicated by the busiest individual processor. We consider torus and mesh topologies, universal fat-trees, and hypercubes; algorithms covered include classical matrix multiplication and direct numerical linear algebra, fast matrix multiplication algorithms, programs that reference arrays, N-body computations, and the FFT. For example, we show that fast matrix multiplication algorithms (e.g., Strassen's) running on a 3D torus will suffer from contention bottlenecks. On the other hand, this network is likely sufficient for a classical matrix multiplication algorithm. Our new lower bounds are matched by existing algorithms only in very few cases, leaving many open problems for network and algorithmic design.
{"title":"Network Topologies and Inevitable Contention","authors":"Grey Ballard, J. Demmel, A. Gearhart, Benjamin Lipshitz, Yishai Oltchik, O. Schwartz, Sivan Toledo","doi":"10.1109/COM-HPC.2016.10","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.10","url":null,"abstract":"Network topologies can have significant effect on the execution costs of parallel algorithms due to inter-processor communication. For particular combinations of computations and network topologies, costly network contention may inevitably become a bottleneck, even if algorithms are optimally designed so that each processor communicates as little as possible. We obtain novel contention lower bounds that are functions of the network and the computation graph parameters. For several combinations of fundamental computations and common network topologies, our new analysis improves upon previous per-processor lower bounds which only specify the number of words communicated by the busiest individual processor. We consider torus and mesh topologies, universal fat-trees, and hypercubes; algorithms covered include classical matrix multiplication and direct numerical linear algebra, fast matrix multiplication algorithms, programs that reference arrays, N-body computations, and the FFT. For example, we show that fast matrix multiplication algorithms (e.g., Strassen's) running on a 3D torus will suffer from contention bottlenecks. On the other hand, this network is likely sufficient for a classical matrix multiplication algorithm. Our new lower bounds are matched by existing algorithms only in very few cases, leaving many open problems for network and algorithmic design.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127557791","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}
François Tessier, Preeti Malakar, V. Vishwanath, E. Jeannot, Florin Isaila
Reading and writing data efficiently from storage systems is critical for high performance data-centric applications. These I/O systems are being increasingly characterized by complex topologies and deeper memory hierarchies. Effective parallel I/O solutions are needed to scale applications on current and future supercomputers. Data aggregation is an efficient approach consisting of electing some processes in charge of aggregating data from a set of neighbors and writing the aggregated data into storage. Thus, the bandwidth use can be optimized while the contention is reduced. In this work, we take into account the network topology for mapping aggregators and we propose an optimized buffering system in order to reduce the aggregation cost. We validate our approach using micro-benchmarks and the I/O kernel of a large-scale cosmology simulation. We show improvements up to 15× faster for I/O operations compared to a standard implementation of MPI I/O.
{"title":"Topology-Aware Data Aggregation for Intensive I/O on Large-Scale Supercomputers","authors":"François Tessier, Preeti Malakar, V. Vishwanath, E. Jeannot, Florin Isaila","doi":"10.1109/COM-HPC.2016.13","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.13","url":null,"abstract":"Reading and writing data efficiently from storage systems is critical for high performance data-centric applications. These I/O systems are being increasingly characterized by complex topologies and deeper memory hierarchies. Effective parallel I/O solutions are needed to scale applications on current and future supercomputers. Data aggregation is an efficient approach consisting of electing some processes in charge of aggregating data from a set of neighbors and writing the aggregated data into storage. Thus, the bandwidth use can be optimized while the contention is reduced. In this work, we take into account the network topology for mapping aggregators and we propose an optimized buffering system in order to reduce the aggregation cost. We validate our approach using micro-benchmarks and the I/O kernel of a large-scale cosmology simulation. We show improvements up to 15× faster for I/O operations compared to a standard implementation of MPI I/O.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127499094","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}
The evolution of massively parallel supercomputers make palpable two issues in particular: the load imbalance and the poor management of data locality in applications. Thus, with the increase of the number of cores and the drastic decrease of amount of memory per core, the large performance needs imply to particularly take care of the load-balancing and as much as possible of the locality of data. One mean to take into account this locality issue relies on the placement of the processing entities and load balancing techniques are relevant in order to improve application performance. With large-scale platforms in mind, we developed a hierarchical and distributed algorithm which aim is to perform a topology-aware load balancing tailored for Charm++ applications. This algorithm is based on both LibTopoMap for the network awareness aspects and on TREEMATCH to determine a relevant placement of the processing entities. We show that the proposed algorithm improves the overall execution time in both the cases of real applications and a synthetic benchmark as well. For this last experiment, we show a scalability up to one millions processing entities.
{"title":"Topology and Affinity Aware Hierarchical and Distributed Load-Balancing in Charm++","authors":"E. Jeannot, Guillaume Mercier, François Tessier","doi":"10.1109/COM-HPC.2016.12","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.12","url":null,"abstract":"The evolution of massively parallel supercomputers make palpable two issues in particular: the load imbalance and the poor management of data locality in applications. Thus, with the increase of the number of cores and the drastic decrease of amount of memory per core, the large performance needs imply to particularly take care of the load-balancing and as much as possible of the locality of data. One mean to take into account this locality issue relies on the placement of the processing entities and load balancing techniques are relevant in order to improve application performance. With large-scale platforms in mind, we developed a hierarchical and distributed algorithm which aim is to perform a topology-aware load balancing tailored for Charm++ applications. This algorithm is based on both LibTopoMap for the network awareness aspects and on TREEMATCH to determine a relevant placement of the processing entities. We show that the proposed algorithm improves the overall execution time in both the cases of real applications and a synthetic benchmark as well. For this last experiment, we show a scalability up to one millions processing entities.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117160004","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}
Ching-Hsiang Chu, Khaled Hamidouche, H. Subramoni, Akshay Venkatesh, B. Elton, D. Panda
Streaming applications, which are data-intensive, have been extensively run on High-Performance Computing (HPC) systems to seek the higher performance and scalability. These applications typically utilize broadcast operations to disseminate in real-time data from a single source to multiple workers, each being a multi-GPU based computing site. State-of-the-art broadcast operations take advantage of InfiniBand (IB) hardware multicast (MCAST) and NVIDIA GPUDirect features to boost inter-node communications performance and scalability. The IB MCAST feature works only with the IB Unreliable Datagram (UD) mechanism and consequently provides unreliable communication for applications. Higher-level libraries and/or runtime environments must handle and provide reliability explicitly. However, handling reliability at that level can be a performance bottleneck for streaming applications. In this paper, we analyze the specific requirements of streaming applications and the performance bottlenecks involved in handling reliability. We show that the traditional Negative-Acknowledgement (NACK) based approach requires the broadcast sender to perform retransmissions for lost packets, degrading streaming throughput. To alleviate this issue, we propose a novel Remote Memory Access (RMA) based scheme to provide high-performance reliability support at the MPI-level. In the proposed scheme, the receivers themselves (as opposed to the sender) retrieve lost packets through RMA operations. Furthermore, we provide an analytical model to illustrate the memory requirements of the proposed RMA-based scheme. Our experimental results show that the proposed scheme introduces nearly no overhead compared to the existing solutions. In a micro-benchmark with injected failures (to simulate unreliable network environments), the proposed scheme shows up to 45% reduction in latency compared to the existing NACK-based scheme. Moreover, with a synthetic streaming benchmark, our design also shows up to a 56% higher broadcast rate compared to the traditional NACK-based scheme on a GPU-dense Cray CS-Storm system with up to 88 NVIDIA K80 GPU cards.
{"title":"Efficient Reliability Support for Hardware Multicast-Based Broadcast in GPU-enabled Streaming Applications","authors":"Ching-Hsiang Chu, Khaled Hamidouche, H. Subramoni, Akshay Venkatesh, B. Elton, D. Panda","doi":"10.1109/COM-HPC.2016.9","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.9","url":null,"abstract":"Streaming applications, which are data-intensive, have been extensively run on High-Performance Computing (HPC) systems to seek the higher performance and scalability. These applications typically utilize broadcast operations to disseminate in real-time data from a single source to multiple workers, each being a multi-GPU based computing site. State-of-the-art broadcast operations take advantage of InfiniBand (IB) hardware multicast (MCAST) and NVIDIA GPUDirect features to boost inter-node communications performance and scalability. The IB MCAST feature works only with the IB Unreliable Datagram (UD) mechanism and consequently provides unreliable communication for applications. Higher-level libraries and/or runtime environments must handle and provide reliability explicitly. However, handling reliability at that level can be a performance bottleneck for streaming applications. In this paper, we analyze the specific requirements of streaming applications and the performance bottlenecks involved in handling reliability. We show that the traditional Negative-Acknowledgement (NACK) based approach requires the broadcast sender to perform retransmissions for lost packets, degrading streaming throughput. To alleviate this issue, we propose a novel Remote Memory Access (RMA) based scheme to provide high-performance reliability support at the MPI-level. In the proposed scheme, the receivers themselves (as opposed to the sender) retrieve lost packets through RMA operations. Furthermore, we provide an analytical model to illustrate the memory requirements of the proposed RMA-based scheme. Our experimental results show that the proposed scheme introduces nearly no overhead compared to the existing solutions. In a micro-benchmark with injected failures (to simulate unreliable network environments), the proposed scheme shows up to 45% reduction in latency compared to the existing NACK-based scheme. Moreover, with a synthetic streaming benchmark, our design also shows up to a 56% higher broadcast rate compared to the traditional NACK-based scheme on a GPU-dense Cray CS-Storm system with up to 88 NVIDIA K80 GPU cards.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125149864","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}
Partitioned Global Address Space (PGAS) parallel programming models can provide an efficient mechanism for managing shared data stored across multiple nodes in a distributed memory system. However, these models are traditionally directly addressed and, for applications with loosely-structured or sparse data, determining the location of a given data element within a PGAS can incur significant overheads. Applications incur additional overhead from the network latency of lookups from remote location resolution structures. Further, for large data, caching such structures locally incurs space and coherence overheads that can limit scaling. We observe that the matching structures used by implementations of the Message Passing Interface (MPI) establish a separation between incoming data writes and the location where data will be stored. In this work, we investigate extending such structures to add a layer of indirection between incoming data reads and the location from which data will be read, effectively extending PGAS models with logical addressing.
{"title":"Extending a Message Passing Runtime to Support Partitioned, Global Logical Address Spaces","authors":"D. B. Larkins, James Dinan","doi":"10.5555/3018058.3018060","DOIUrl":"https://doi.org/10.5555/3018058.3018060","url":null,"abstract":"Partitioned Global Address Space (PGAS) parallel programming models can provide an efficient mechanism for managing shared data stored across multiple nodes in a distributed memory system. However, these models are traditionally directly addressed and, for applications with loosely-structured or sparse data, determining the location of a given data element within a PGAS can incur significant overheads. Applications incur additional overhead from the network latency of lookups from remote location resolution structures. Further, for large data, caching such structures locally incurs space and coherence overheads that can limit scaling. We observe that the matching structures used by implementations of the Message Passing Interface (MPI) establish a separation between incoming data writes and the location where data will be stored. In this work, we investigate extending such structures to add a layer of indirection between incoming data reads and the location from which data will be read, effectively extending PGAS models with logical addressing.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128806517","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}
Huansong Fu, S. Pophale, Manjunath Gorentla Venkata, Weikuan Yu
Despite the popularity of MPI for high performance computing, the startup of MPI programs faces a scalability challenge as both the execution time and memory consumption increase drastically at scale. We have examined this problem using the collective modules of Cheetah and Tuned in Open MPI as representative implementations. Previous improvements for collectives have focused on algorithmic advances and hardware off-load. In this paper, we examine the startup cost of the collective module within a communicator and explore various techniques to improve its efficiency and scalability. Accordingly, we have developed a new scalable startup scheme with three internal techniques, namely Delayed Initialization, Module Sharing and Prediction-based Topology Setup (DISP). Our DISP scheme greatly benefits the collective initialization of the Cheetah module. At the same time, it helps boost the performance of non-collective initialization in the Tuned module. We evaluate the performance of our implementation on Titan supercomputer at ORNL with up to 4096 processes. The results show that our delayed initialization can speed up the startup of Tuned and Cheetah by an average of 32.0% and 29.2%, respectively, our module sharing can reduce the memory consumption of Tuned and Cheetah by up to 24.1% and 83.5%, respectively, and our prediction-based topology setup can speed up the startup of Cheetah by up to 80%.
尽管MPI在高性能计算方面很受欢迎,但MPI程序的启动面临着可伸缩性的挑战,因为执行时间和内存消耗都在大规模地急剧增加。我们使用Cheetah和Tuned in Open MPI的集合模块作为代表性实现来研究这个问题。之前对集体的改进主要集中在算法进步和硬件卸载上。在本文中,我们研究了通信器中集合模块的启动成本,并探索了各种技术来提高其效率和可扩展性。因此,我们开发了一种新的可扩展启动方案,其中包含三种内部技术,即延迟初始化,模块共享和基于预测的拓扑设置(DISP)。我们的DISP方案极大地有利于Cheetah模块的集体初始化。同时,它有助于提高Tuned模块中非集体初始化的性能。我们在ORNL的Titan超级计算机上使用多达4096个进程来评估我们的实现性能。结果表明,延迟初始化可以使tuning和Cheetah的启动速度平均分别提高32.0%和29.2%,模块共享可以使tuning和Cheetah的内存消耗分别降低24.1%和83.5%,基于预测的拓扑设置可以使Cheetah的启动速度提高80%。
{"title":"DISP: Optimizations towards Scalable MPI Startup","authors":"Huansong Fu, S. Pophale, Manjunath Gorentla Venkata, Weikuan Yu","doi":"10.1109/COM-HPC.2016.11","DOIUrl":"https://doi.org/10.1109/COM-HPC.2016.11","url":null,"abstract":"Despite the popularity of MPI for high performance computing, the startup of MPI programs faces a scalability challenge as both the execution time and memory consumption increase drastically at scale. We have examined this problem using the collective modules of Cheetah and Tuned in Open MPI as representative implementations. Previous improvements for collectives have focused on algorithmic advances and hardware off-load. In this paper, we examine the startup cost of the collective module within a communicator and explore various techniques to improve its efficiency and scalability. Accordingly, we have developed a new scalable startup scheme with three internal techniques, namely Delayed Initialization, Module Sharing and Prediction-based Topology Setup (DISP). Our DISP scheme greatly benefits the collective initialization of the Cheetah module. At the same time, it helps boost the performance of non-collective initialization in the Tuned module. We evaluate the performance of our implementation on Titan supercomputer at ORNL with up to 4096 processes. The results show that our delayed initialization can speed up the startup of Tuned and Cheetah by an average of 32.0% and 29.2%, respectively, our module sharing can reduce the memory consumption of Tuned and Cheetah by up to 24.1% and 83.5%, respectively, and our prediction-based topology setup can speed up the startup of Cheetah by up to 80%.","PeriodicalId":332852,"journal":{"name":"2016 First International Workshop on Communication Optimizations in HPC (COMHPC)","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2016-11-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116594703","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}