MORITZ, C. A., AND FRANK, M. I. LoGPC: Modeling Network Contention in MessagePassing Programs. In Proceedings of the 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....model to take into account other factors that influence application performance, and tailor the model for different communication layers (e.g. MPI) and architectures. Alexandrov et al. extend it with G (LogGP [1] to capture the cost of large messages. Moritz and Frank further extend LogGP (LoGPC [17]) to take into account the effects of message pipelining and network contention. AlTawil and Moritz [16] use LogGP to analyze the behavior of MPI under the different send protocols mandated by the standard. Ino et al. tailor the LogGP model (LogGPS [12] to account for the synchronization costs ....

C. A. Moritz and M. Frank. LoGPC: Modeling network contention in message-passing programs. In Measurement and Modeling of Computer Systems, pages 254--263, 1998. 10

....the goal of high performance computing is to minimise the response time rather than utilising processors or increasing throughput [1] we have con centrated both on utilisation side and response time. In the literature, there are a number of cost models such as the ones defined in [1] 2] 3] 4] [5] and [6] We defined scheduling and data partitioning schemes that can work together. The parameters for defined schemes reflect the features of the chosen parallel system architecture and algorithm class. The defined schemes could be used for finding the optimal number of processors and ....

Moritz C A, Frank M, LoGPC: Modeling Network Contention in Message-Passing Programs, ACM Joint International Conference on Measurement and Modeling of Computer Systems, ACM Sigmetrics/Performance 98, Wisconsin, June 1998.

....model to take into account other factors that influence application performance, and tailor the model for different communication layers (e.g. MPI) and architectures. Alexandrov et al. extend it with G (LogGP [1] to capture the cost of large messages. Moritz and Frank further extend LogGP (LoGPC [16]) to take into account the effects of message pipelining and network contention. Al Tawil and Moritz [15] use LogGP to analyze the behavior of MPI under the different send protocols mandated by the standard. Ino et al. tailor the LogGP model (LogGPS [11] to account for the synchronization costs ....

C. A. Moritz and M. Frank. LoGPC: Modeling network contention in message-passing programs. In Measurement and Modeling of Computer Systems, pages 254--263, 1998.

....not. In our research this is not a problem, as we only apply communication patterns that are expected to perform well on most network topologies used today. Still, because P 3PC is similar to the LogGP model, it can easily be extended to account for contention, in the same manner as described in [2]. As the P 3PC model stresses the importance of benchmarking to obtain accurate values for the model parameters, one may argue that the predictive power of the model is limited. However, the model does not specifically enforce a large number of measurements to be performed. As for models that ....

C. Andras Moritz and M. Frank. LoGPC: Modeling Network Contention in Message-Passing Programs. IEEE Transactions on Parallel and Distributed Systems, 12(4):404--415, Apr. 2001.

....the knee of the curve yet no real justification or evidence is given to support this statement. In between the basic PRAM and the very specific network models, LogP is a single point among several other models with more or less detail (the BSP[6] model is less detailed, LogGP[1] and LoGPC[4] are more detailed) 2.4 Tractability of Analysis It is unclear how tractable the analysis of complex algorithms is under the LogP model. In Section 6.5, this question is raised, but is not satisfactorily addressed. It is claimed that in some algorithms or some machines, certain parameters (out ....

Csaba Andras Moritz and Matthew I. Frank. LoGPC: Modeling network contention in message-passing programs. ACM SIGMETRICS, 1998.

....sufficient bandwidth. 6 Discussion and Conclusions The most common approach to achieving high performance communication is to tailor the communication patterns and algorithms to the properties of the underlying network. This is the basis for programming models like LogP and its derivatives [11, 1, 15, 22] and the Postal 15 all to all 1 2 4 5 6 bisection 384 1.5K 6.1K 24.5K message size 1000 1500 2000 2500 3000 3500 4000 4500 aggregate bandwidth [Mb s] all to all phases 1 2 4 5 6 bisection 384 1.5K 6.1K 24.5K message size 500 1000 1500 2000 2500 3000 3500 ....

C. A. Moritz and M. I. Frank, "LoGPC: modeling network contention in message-passing programs". In SIGMETRICS Conf. Measurement & Modeling of Comput. Syst., pp. 254--263, Jun 1998.

....Figure 2) Because network contention affects the transfer time of data, it can severely lengthen the execution time of jobs. Both I O traffic and intra job communication can suffer from network contention. Most previous research has focused on network contention due to intra job communication [30, 4, 8, 29, 27, 24], whereas our focus is on contention due to parallel I O traffic. Parallel I O traffic bears many similarities to intra job communication; yet it differs from intra job communication, both in the size of the messages, and also in what links are stressed in the network. Especially since the I O ....

Csaba A. Moritz and Matthew I. Frank. LoGPC: Modeling network contention in message-passing programs. In Proceedings of Sigmetrics, 1998.

....chip. 4. Preliminary Experimental Evaluation This section presents our preliminary findings regarding possible performance improvements due to improved network utilization on a BT enabled Raw system. In our evaluation we used two analytical frameworks, the SimpleFit model [9] and the LoGPC model [10]. Our earlier studies presented SimpleFit [9] a novel analytical framework that designers can use to reason about the design space of RAW microprocessors in a billion transistor era. This model is also generalizable to other single chip systems. Although the optimal machine configurations ....

....communication patterns or those using large messages, network contention and contention for message processing resources can be a significant part of the total execution time. Our earlier studies on modeling network contention in message passing programs presented a new cost model called LoGPC [10]. Based on that model we expect network contention to be more significant because of increased average distance traversed by a message. 0 500 1000 1500 2000 2500 Jacobi Matmul FFT Nbody LCS Processors N=10 2 N=10 4 N=10 6 N=10 8 Figure 6: Number of Processors in optimal machine ....

[Article contains additional citation context not shown here]

Csaba Andras Moritz and Matthew I. Frank, "LoGPC: Modeling Network Contention in Message-Passing Programs," IEEE Transactions on Parallel and Distributed Systems, Vol. 12, No. 4, April 2001.

....on distributed memory machines. An extension of LogP for large messages is presented in [3] and a new parameter G is introduced: the Gap per byte or the time per byte for long messages. The reciprocal of G characterizes the available per processor communication bandwidth for long messages. LoGPC [5] is a new model where application specific parameters are introduced to account for network and resource contention effects. LogP is quantified for low overhead local area networks in [14] The performance assessment of LogP for fast network interfaces is presented in [11] The majority of ....

C Andras Moritz, Matthew I.Frank. "LoGPC: Modeling Network Contention in Message-Passing Applications." In Performance Evaluation Review , Special Issue Volume 26 no. 1 and Proceedings of ACM Joint International Conference on Measurement and Modeling of Computer Systems, SIGMETRICS /PERFORMANCE 98, Wisconsin Madison, June 1998.

No context found.

MORITZ, C. A., AND FRANK, M. I. LoGPC: Modeling Network Contention in MessagePassing Programs. In Proceedings of the 1998.

No context found.

MORITZ, C. A., AND FRANK, M. I. LoGPC: Modeling Network Contention in MessagePassing Programs. In Proceedings of the 1998.

No context found.

C. Andras Moritz and M.I. Frank, "LoGPC: Modeling Network Contention in Message-Passing Programs," IEEE Trans. Parallel and Distributed Systems, vol. 12, no. 4, pp. 404-415, Apr. 2001.

No context found.

C. Andras Moritz and M.I. Frank. LoGPC: Modeling Network Contention in MessagePassing Programs. IEEE Transactions on Parallel and Distributed Systems, 12(4):404-- 415, April 2001.

No context found.

C. Andras Moritz and M. Frank. LoGPC: Modeling Network Contention in Message-Passing Programs. IEEE Transactions on Parallel and Distributed Systems, 12(4):404--415, Apr. 2001.

No context found.

C. Andras Moritz and M.I. Frank. LoGPC: Modeling Network Contention in Message-Passing Programs. IEEE Transactions on Parallel and Distributed Systems, 12(4):404-415, April 2001.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC