Advertisement

Modeling Contention and Mapping Effects in Multi-core Clusters

  • Juan-Antonio Rico-Gallego
  • Juan-Carlos Díaz-Martín
  • Alexey L. Lastovetsky
Conference paper
Part of the Lecture Notes in Computer Science book series (LNCS, volume 9523)

Abstract

Modeling and formal analysis of parallel algorithms contribute to optimize their performance. Modern multi-core are complex machines composed of heterogeneous shared communication channels. Parallel Performance Models estimate the cost and capture the behavior of parallel algorithms through a set of parameters, providing valuable information about the behavior of the algorithm in these platforms. LogGP is a representative model using network related parameters to predict the cost of parallel algorithms as a sequence of point-to-point transmissions. Although extensions have been proposed for covering issues derived from modern platforms complexities as contention and channels hierarchy, such specific extensions are not enough to meaningfully and accurately model more than simple algorithms. \(\uptau \)–Lop is an alternative model that takes as a building block for modeling parallel algorithms the concept of concurrent transfers, that helps to capture algorithms behavior and allows to represent and accurately predict their cost in multi-core clusters. This paper shows the analysis capabilities of \(\uptau \)–Lop through two cases of study involving elaborated MPI collective operations.

Notes

Acknowledgments

This work has been partially supported by the computing facilities of the Extremadura Supercomputing Center (CenitS), by EU under the COST programme Action IC1305, ‘Network for Sustainable Ultrascale Computing (NESUS)’, and the Extremadura Government through the FEDER Funds.

References

  1. 1.
    Alexandrov, A., Lonescu, M.F., Schauser, K.E., Scheiman, C.: Loggp: incorporating long messages into the logp model - one step closer towards a realistic model for parallel computation. In: Proceedings of the Seventh Annual ACM Symposium on Parallel Algorithms and Architectures, pp. 95–105. SPAA 1995, NY, USA (1995)Google Scholar
  2. 2.
    Cameron, K.W., Ge, R., Sun, X.H.: \(\log _{m} p\) and \(\log _{3} p\): accurate analytical models of point-to-point communication in distributed systems. IEEE Trans. Comput. 56(3), 314–327 (2007)CrossRefMathSciNetGoogle Scholar
  3. 3.
    Ino, F., Fujimoto, N., Hagihara, K.: Loggps: a parallel computational model for synchronization analysis. SIGPLAN Not. 36, pp. 133–142 (2001). http://doi.acm.org/10.1145/568014.379592
  4. 4.
    Kielmann, T., Bal, H.E., Verstoep, K.: Fast measurement of LogP parameters for message passing platforms. In: Rolim, J.D.P. (ed.) IPDPS-WS 2000. LNCS, vol. 1800, pp. 1176–1183. Springer, Heidelberg (2000) CrossRefGoogle Scholar
  5. 5.
    Lastovetsky, A., Rychkov, V.: Accurate and efficient estimation of parameters of heterogeneous communication performance models. Int. J. High Perform. Comput. Appl. 23(2), 123–139 (2009). http://dx.doi.org/10.1177/1094342009103947 CrossRefGoogle Scholar
  6. 6.
    Martinasso, M., Méhaut, J.-F.: Prediction of communication latency over complex network behaviors on SMP clusters. In: Bravetti, M., Kloul, L., Zavattaro, G. (eds.) EPEW/WS-EM 2005. LNCS, vol. 3670, pp. 172–186. Springer, Heidelberg (2005) CrossRefGoogle Scholar
  7. 7.
    Moritz, C.A., Frank, M.I.: Logpc: modeling network contention in message-passing programs. IEEE Trans. Parallel Distrib. Syst. 12(4), 404–415 (2001). http://dx.doi.org/10.1109/71.920589 CrossRefGoogle Scholar
  8. 8.
    MPI Forum: \({\sf MPI}\): A message-passing interface standard. Version 3.0 (September 21th 2012). http://www.mpi-forum.org
  9. 9.
    Pješivac-Grbović, J., Angskun, T., Bosilca, G., Fagg, G.E., Gabriel, E., Dongarra, J.J.: Performance analysis of MPI collective operations. Cluster Comput. 10, 127–143 (2007). http://dl.acm.org/citation.cfm?id=1265235.1265248 CrossRefGoogle Scholar
  10. 10.
    Rico-Gallego, J.A., Díaz-Martín, J.C.: \(\tau \)-lop: modeling performance of shared memory \({ MPI}\). Parallel Comput. 46, 14–31 (2015)CrossRefGoogle Scholar
  11. 11.
    Steffenel, L.: Modeling network contention effects on all-to-all operations. In: 2006 IEEE International Conference on Cluster Computing, pp. 1–10 (2006)Google Scholar
  12. 12.
    Yuan, L., Zhang, Y., Tang, Y., Rao, L., Sun, X.: Loggph: a parallel computational model with hierarchical communication awareness. In: 2010 IEEE 13th International Conference on Computational Science and Engineering (CSE), vol. 2, pp. 268–274 (2010)Google Scholar

Copyright information

© Springer International Publishing Switzerland 2015

Authors and Affiliations

  • Juan-Antonio Rico-Gallego
    • 1
  • Juan-Carlos Díaz-Martín
    • 1
  • Alexey L. Lastovetsky
    • 2
  1. 1.University of ExtremaduraCáceresSpain
  2. 2.University College DublinBelfieldIreland

Personalised recommendations