PARALLEL DATA LAB

I/O Workload Characterization: Data Mining meets Traffic Modeling

Traffic modeling of storage workloads is extremely helpful in evaluating system designs. The work involves the following two aspects. The first is to discover and to quantify the most important features of the traffic data. Two example features are temporal burstiness and spatial locality. In addition, it's even harder to determine how these features affect the performance of the traffic data in real systems. Secondly, we need an efficient statistical model to generate synthetic workloads of similar behavior as the real ones. Traditional models such as Poisson are inadequate in generating timestamps for traffic data of strong burstiness, not mentioning generating multi-dimensional traffic.

This project is to solve the above problem. Our previous work has focused on the spatio-temporal behavior of traffic data, more specifically, the temporal burstiness and spatial locality of I/O workload. Our proposed tool, entropy plot, is able to quantify the temporal burstiness and spatial locality in traffic data. The B-model generates the timestamps for the synthetic traffic to imitate the temporal burstiness of real traffic data. The PQRS model goes one step further by generating both the timestamps and request locations for synthetic traces. The ongoing work is to augment the model to deal with more dimensionality.

 



2- and 3-dimensional representations of real traffic data
showing burstiness along time and space.


People

FACULTY

Anastassia Ailamaki
Christos Faloutsos

STUDENTS

Mengzhi Wang


Publications

  • Storage Device Performance Prediction with CART Models. Mengzhi Wang, Kinman Au, Anastassia Ailamaki, Anthony Brockwell, Christos Faloutsos, and Gregory R. Ganger. Proc. 12th Annual Meeting of the IEEE/ACM International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems (MASCOTS). Volendam, The Netherlands. October 5-7, 2004. Supercedes Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-04-103, March 2004.
    Abstract / Postscript [908K] / PDF [122K]

  • Storage Device Performance Prediction with CART Models [Extended Abstract]. Mengzhi Wang, Kinman Au, Anastassia Ailamaki, Anthony Brockwell, Christos Faloutsos, and Gregory R. Ganger. Proceedings: Poster Session. Joint International Conference on Measurement and Modeling of Computer Systems. ACM SIGMETRICS/Performance 2004. June 12th-16th 2004, Columbia University, New York.
    Abstract / Postscript [400K] / PDF [64K]

  • SIMFLEX: A Fast, Accurate, Flexible Full-System Simulation Framework for Performance Evaluation of Server Architecture. Nikolaos Hardavellas, Stephen Somogyi, Thomas F. Wenisch, Roland E. Wunderlich, Shelley Chen, Jangwoo Kim, Babak Falsafi, James C. Hoe, and Andreas G. Nowatzyk. ACM SIGMETRICS Performance Evaluation Review (PER) Special Issue on Tools for Computer Architecture Research, Volume 31, Number 4, pages 31-35, March 2004.
    Abstract / PDF [96K]

  • Capturing the Spatio-Temporal Behavior of Real Traffic Data. Mengzhi Wang, Anastassia Ailamaki, and Christos Faloutsos. Performance 2002, September, 2002, Rome, Italy. Best Student Paper Award.
    Abstract / PDF [1.9M]

  • Data Mining Meets Performance Evaluation: Fast Algorithms for Modeling Bursty Traffic. M. Wang, T. Madhyastha, N.H. Chan, S. Papadimitriou, C. Faloutsos. 18th International Conference on Data Engineering, February 26-March 1, 2002 San Jose, California. Also available as a technical report CMU-CS-01-101.
    Abstract / Postscript [2.25M] / PDF [358K]


Acknowledgements

We thank the members and companies of the PDL Consortium: Amazon, Google, Hitachi Ltd., Honda, Intel Corporation, IBM, Meta, Microsoft Research, Oracle Corporation, Pure Storage, Salesforce, Samsung Semiconductor Inc., Two Sigma, and Western Digital for their interest, insights, feedback, and support.