MPI Collectives on Modern Multicore Clusters: Performance Optimizations and Communication Characteristics

Mamidala, Amith R.; Kumar, Rahul; D, de Lara; Panda, D.K.

doi:10.1109/ccgrid.2008.87

Cited by 70 publications

(27 citation statements)

References 5 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…This stems from processing overhead within MPI on that core for collective operations, in particular from large message broadcasts. As additional experiments verified, this core acts as a local leader for optimized collectives in MPI, which first gather on-node data to the leader and then use this buffer for broadcast [21].…”

Section: A Performance Data In the Application Domain (H → A)mentioning

confidence: 95%

Interpreting Performance Data across Intuitive Domains

Schulz

Levine

Bremer

et al. 2011

2011 International Conference on Parallel Processing

View full text Add to dashboard Cite

Abstract-To exploit the capabilities of current and future systems, developers must understand the interplay between onnode performance, domain decomposition, and an application's intrinsic communication patterns. While tools exist to gather and analyze data for each of these components individually, the resulting information is generally processed in isolation and presented in an abstract, categorical fashion unintuitive to most users. In this paper we present the HAC model, in which we identify the three domains of performance data most familiar to the user: (i) the application domain containing the application's working set, (ii) the hardware domain of the compute and network devices, and (iii) the communication domain of logical data transfers.We show that taking data from each of these domains and projecting, visualizing, and correlating it to the other domains can give valuable insights into the behavior of parallel application codes. The HAC abstraction opens the door for a new generation of tools that can help users more easily and intuitively associate performance data with root causes in the hardware system, the application's structure, and in its communication behavior, and by doing so leads to an improved understanding of the performance of their codes. Keywords-Performance Analysis and Visualization I. MOTIVATIONBy the end of this decade, exascale machines are expected to support 500 million to 4 billion concurrent tasks. At the same time the complexity of both architectures and applications will increase significantly, leading to often unexpected and unpredictable interactions between applications, middleware, and underlying hardware. In order to exploit such machines fully, programmers must be able to measure, analyze, and understand the behavior of their applications and its corresponding impact on the overall performance in great detail.Most current tools gather and present performance data in the native hardware or MPI rank domain, relying on the user to infer causal relationships and interdependencies. However, neither of these two domains is intuitive or familiar to the users. As a result, data appear as abstract or highly discrete, and domain and performance experts must expect

show abstract

Section: A Performance Data In the Application Domain (H → A)mentioning

confidence: 95%

Interpreting Performance Data across Intuitive Domains

Schulz

Levine

Bremer

et al. 2011

2011 International Conference on Parallel Processing

View full text Add to dashboard Cite

show abstract

“…Much work has been done to improve MPI performance on SMP-CMP clusters. Techniques were developed to improve both point-to-point communications [7], [9], [10], [12] and collective communications [14], [15], [17], [19]. These optimizations further differentiate the communication performance in the multi-layer communication infrastructure in SMP-CMP clusters and manifest the impacts of processor affinity.…”

Section: Related Workmentioning

confidence: 99%

Processor affinity and MPI performance on SMP-CMP clusters

Zhang

Yuan

Srinivasan

2010

2010 IEEE International Symposium on Parallel &Amp; Distributed Processing, Workshops and PHD Forum (IPDPSW)

View full text Add to dashboard Cite

Abstract-Clusters of Symmetric MultiProcessing (SMP) nodes with multi-core Chip-Multiprocessors (CMP), also known as SMP-CMP clusters, are becoming ubiquitous today. For Message Passing interface (MPI) programs, such clusters have a multilayer hierarchical communication structure: the performance of intra-node communication is usually higher than that of internode communication; and the performance of intra-node communication is not uniform with communications between cores within a chip offering higher performance than communications between cores in different chips. As a result, the mapping from Message Passing Interface (MPI) processes to cores within each compute node, that is, processor affinity, may significantly affect the performance of intra-node communication, which in turn may impact the overall performance of MPI applications. In this work, we study the impacts of processor affinity on MPI performance in SMP-CMP clusters through extensive benchmarking and identify the conditions when processor affinity is (or is not) a major factor that affects performance.

show abstract

“…On modern manycore clusters, this aim is supported by improvement on the intra-node side of collective operations which benefits from the sharedmemory between local processes [4,11]. Intranode optimizations are now often combined with inter-node communication within hierarchical algorithms.…”

Section: Collective Operations On Many-core Clustersmentioning

confidence: 99%