Realistic communication model for parallel computing on cluster

Measurement and Prediction of Communication Delays in Myrinet Networks

“…They examined methods for empirically obtaining the parameters used in the LogGP model, and then showed the effects of each of those parameters on the performance of programs in a parallel benchmark suite. Reference [7] introduced another extension to the LogP model, targeted for cluster computing, using 8 parameters Although of significant theoretical importance, the LogP, LogGP, and other extended models are not directly comparable to the results of this paper since the focus is primarily on measurement and prediction, rather than modeling, of communication delays. The work in this paper can be considered as a more accurate method for predicting the L+o (network latency and overhead) portion of the LogP model.…”

An EasyGrid portal for scheduling system‐aware applications on computational Grids

“…The LogGP model [8] uses an additional parameter to represent the bandwidth when sending long messages. Tam and Wang [9] proposed a model whose architectural parameters are based on those of the LogGP model, but are expressed as functions of message length rather than as constants. LogP and its extensions have successfully been used to model a number of parallel machines, including Berkeley NOW, CM-5, Dash, IBM SP1 and SP2, Intel Paragon and computing clusters [7][8][9][10][11].…”

“…As such, while it is not required to be as detailed as [9], for example, it needs to be sufficiently precise to influence scheduling decisions correctly. Unfortunately, this model is sufficiently different from the models adopted by the scheduling community, which makes it difficult to apply traditional scheduling algorithm design methods (e.g.…”

PEMPIs: A New Methodology for Modeling and Prediction of MPI Programs Performance

“…According to the TOP500 list, (1) 80% of the 500 most performing available machines can be classified as a cluster of computers. 3 The main selling point is their cost-effectiveness as compared to conventional parallel machines architectures. Their success is also based on recent advances in commodity hardware, e.g.…”

“…(2) The development of applications for clusters of computers faces the same dilemma as traditional parallel computing-to design high-level, architectural independent algorithms which will execute efficiently on a wide range of current and future cluster platforms. (3) Many points should be considered by the application developer such as load imbalance, insufficient parallelism, synchronization loss, communication loss, and resource contention. (2) In addition to the difficulties in the application development, there is a lack of sufficiently accurate and easy-to-use performance analysis and prediction methodologies for message passing programs.…”