Evaluation of Remote Memory Access Communication on the Cray XT3

Tipparaju, Vinod; Kot, Andriy; Nieplocha, Jarek; Bruggencate, Monika ten; Chrisochoides, Nikos

doi:10.1109/ipdps.2007.370478

Cited by 13 publications

(10 citation statements)

References 8 publications

(6 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The reason for choosing MG is the same as provided in [23]. The major communication work is done using point-to-point communication, and it is relatively easy to convert using the GET/PUT protocol.…”

Section: Nas Mg Benchmarkmentioning

confidence: 99%

“…The GET/PUT protocol has been the subject of a lot of attention in an effort to achieve efficient parallel programming [15,26,25,23], since it can eliminate the implicit synchronization that is difficult to avoid in MPI message communication. The initiators of onesided communication, GP-SHMEM [19] for example, trigger GET or PUT operation, whereas the target of the operation has no means of knowing when the one-sided communication takes place.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Audit: A new synchronization API for the GET/PUT protocol

Hori

Lee

Sato

2012

Journal of Parallel and Distributed Computing

View full text Add to dashboard Cite

Section: Nas Mg Benchmarkmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Audit: A new synchronization API for the GET/PUT protocol

Hori

Lee

Sato

2012

Journal of Parallel and Distributed Computing

View full text Add to dashboard Cite

“…In addition, to reduce the message communication latency, highend systems like BlueGene/L [2], BlueGene/P [3], or Cray XT [4] implement Remote Memory Access (RMA) [5]. RMA allows a given board to directly access the memory of another board.…”

Section: Introductionmentioning

confidence: 99%

A cost-effective heuristic to schedule local and remote memory in cluster computers

et al. 2011

View full text Add to dashboard Cite

Cluster computers represent a cost-effective alternative solution to supercomputers. In these systems, it is common to constrain the memory address space of a given processor to the local motherboard. Constraining the system in this way is much cheaper than using a full-fledged shared memory implementation among motherboards. However, memory usage among motherboards can be unfairly balanced.On the other hand, remote memory access (RMA) hardware provides fast interconnects among the motherboards of a cluster. RMA devices can be used to access remote RAM memory from a local motherboard. This work focuses on this capability in order to achieve a better global use of the total RAM memory in the system. More precisely, the address space of local applications is extended to remote motherboards and is used to access remote RAM memory. This paper presents an ideal memory scheduling algorithm and proposes a cost-effective heuristic to allocate local and remote memory among local applications. Compared to the devised ideal algorithm, the heuristic obtains the same (or closely resembling) results while largely reducing the computational cost. In addition, we analyze the impact on the performance of stand alone applications varying the memory distribution among regions (local, local to board, and remote). Then, this study is extended to any number of concurrent applications. Experimental results show that a QoS parameter is needed in order to avoid unacceptable performance degradation.

show abstract

“…So far, Remote Memory Access (RMA) hardware [2], which allows a given node to directly access remote memory, has been only available in supercomputer systems like BlueGene/L [3], BlueGene/P [4], or Cray XT [5]. Nevertheless, commodity implementations for cluster computers are already entering the market.…”

Section: Introductionmentioning

confidence: 99%

A Cluster Computer Performance Predictor for Memory Scheduling

Serrano

Sahuquillo

Hassan

et al. 2011

Algorithms and Architectures for Parallel Processing

View full text Add to dashboard Cite

Abstract. Remote Memory Access (RMA) hardware allow a given motherboard in a cluster to directly access the memory installed in a remote motherboard of the same cluster. In recent works, this characteristic has been used to extend the addressable memory space of selected motherboards, which enable a better balance of main memory resources among cluster applications. This way is much more cost-effective than than implementing a full-fledged shared memory system. In this context, the memory scheduler is in charge of finding a suitable distribution of local and remote memory that maximizes the performance and guarantees a minimum QoS among the applications. Note that since changing the memory distribution is a slow process involving several motherboards, the memory scheduler needs to make sure that the target distribution provides better performance than the current one. In this paper, a performance predictor is designed in order to find the best memory distribution for a given set of applications executing in a cluster motherboard. The predictor uses simple hardware counters to estimate the expected impact on performance of the different memory distributions. The hardware counters provide the predictor with the information about the time spent in processor, memory access and network. The performance model used by the predictor has been validated in a detailed microarchitectural simulator using real benchmarks. Results show that the prediction accuracy never deviates more than 5% compared to the real results, being less than 0.5% in most of the cases.

show abstract

Evaluation of Remote Memory Access Communication on the Cray XT3

Cited by 13 publications

References 8 publications

Audit: A new synchronization API for the GET/PUT protocol

Audit: A new synchronization API for the GET/PUT protocol

A cost-effective heuristic to schedule local and remote memory in cluster computers

A Cluster Computer Performance Predictor for Memory Scheduling

Contact Info

Product

Resources

About