REED, D. A., NICKOLAYEV, O. Y., AND ROTH, P. C. Real-time statistical clustering for event trace reduction. Journal of Supercomputing Applications and High-Performance Computing 11, 2 (1997), 144--159.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....with detailed knowledge about the architecture to detect performance bottlenecks. Such a system should be able to capture data from all relevant sources while incurring minimal overhead on the application. It should be flexible enough to support advanced data capturing and data analysis techniques [10, 8]. Finally, it should be scalable up to the maximum size Origin system. The goal of this work is to design and implement a low overhead, tracebased scalable performance data gathering system. The evaluation work will investigate the total overhead and impact on the application and the tradeoffs in ....

Nickolayev, O. Y., Roth, P. C., and Reed, D.A., Real-Time Statistical Clustering for Event Trace Reduction.

....of mean network latency from machine at Urbana Illinois. 5] applied statistical sampling to estimate power demands in circuits, and compared the results from stratified sampling to those from simple random sampling experiments. In the area of performance analysis, statistical clustering [9] and projection pursuit [12] have proved to be efficient techniques to handle large amounts of performance data. Clustering can reduce the number of processors to be monitored, and projection pursuit can reduce the number of captured metrics. These techniques, however, require periodic analysis of ....

REED, D. A., NICKOLAYEV, O. Y., AND ROTH, P. C. Real-time statistical clustering for event trace reduction. Journal of Supercomputing Applications and High-Performance Computing 11, 2 (1997), 144--159.

....to more equivalence classes. We believe this type of behavior leading to a set of clusters is common for many types of grid applications. Standard clustering techniques can be used to identify equivalence classes of points, yielding cluster centroids and variance factors in the N dimensions [17]. Similarly, equivalence classes can also be identified in the performance metric space. When such equivalence classes exist, contract validation can be thought of as determining whether the metrics measured at runtime fall within an acceptable range of a cluster centroid. The definition of the ....

REED, D. A., NICKOLAYEV, O. Y., AND ROTH, P. C. Real-Time Statistical Clustering for Event Trace Reduction. Journal of Supercomputing Applications and High-Performance Computing 11, 2 (1997), 144 159.

....behavior. After capturing traces from each code on each system configuration, we smoothed the trace data via moving window averages. We then normalized the smoothed data before input to the polyline fitting algorithm. This smoothing and normalization simplified comparison across platforms [7]. 5.1 CSAR Rocket The CSAR Rocket Code [12] is an integrated, wholesystem solid propellant rocket simulation. It focuses on propellant combustion, turbulent flows, rocket case and nozzle structures under both normal and abnormal operations, component aging and other failure models. We chose to ....

....have only characterized the behavior of a single node on a parallel system. Our approach can be easily extended to all nodes by creating a signature for each. By comparing signatures across nodes, we found that behavior of the nodes can be usually divided into several equivalence classes, as in [7]. Hence, only one signature is required for each equivalence class, saving much storage space and computation overhead. For example, the CSAR Rocket s node 0 is the only one that performs disk writes. Similarly, all nodes of the Cactus Wavetoy code have homogeneous disk write patterns (99 ....

Reed, D. A., Nickolayev, O. Y., and Roth, P. C. Real-Time Statistical Clustering for Event Trace Reduction, Journal of Supercomputing Applications and High-Performance Computing, Summer 1997, Vol. 11, No. 2, pp. 144-159.

....each measured on a set of P parallel tasks. Conceptually, one can then view an event trace asdefi a set of n dynamic performance metrics, m i (t) on eachofP tasks (m 1 (t)#m 2 (t)#: #m n (t) p p 2 [1: P ] that describe parallel system characteristics as a function of time t. Following [26], if R i denotes the range of metric m i (t) we call the Cartesian product M = R 1 Theta R 2 Theta : Theta Rn a performance metric space. Thus, the ordered n tuples (m 1 (t) 2 R 1 # m 2 (t) 2 R 2 # : # mn (t) 2 Rn ) 1) are points in M (t) and the event tracedefi the temporal evolution ....

....each processor in the cluster trace similar trajectories. By periodically computing cluster membership using performance metrics from each processor, an event tracing system can capture and extract traces only from representative members of each cluster, dramatically reducing the total data volume [26]. As an example of the possible data reduction due to statistical clustering, consider an SPMD code that relies on a master task to read initialization data and allocate worktoasetofN worker tasks. If the behavior of all workers is similar, clustering identifi two clusters, one with cardinality ....

[Article contains additional citation context not shown here]

Nickolayev, O. Y., Roth, P. C., and Reed, D. A. Real-time Statistical Clustering for Event Trace Reduction. International Journal of Supercomputer Applications and High Performance Computing (1997).

....each measured on a set of P parallel tasks. Conceptually, one can then view an event trace as defining a set of n dynamic performance metrics, m i (t) on each of P tasks (m 1 (t) m 2 (t) mn (t) p p 2 [1: P ] that describe parallel system characteristics as a function of time t. Following [26], if R i denotes the range of metric m i (t) we call the Cartesian product M = R 1 Theta R 2 Theta : Theta Rn a performance metric space. Thus, the ordered n tuples (m 1 (t) 2 R 1 ; m 2 (t) 2 R 2 ; mn (t) 2 Rn ) 1) are points in M (t) and the event trace defines the temporal ....

....each processor in the cluster trace similar trajectories. By periodically computing cluster membership using performance metrics from each processor, an event tracing system can capture and extract traces only from representative members of each cluster, dramatically reducing the total data volume [26]. As an example of the possible data reduction due to statistical clustering, consider an SPMD code that relies on a master task to read initialization data and allocate work to a set of N worker tasks. If the behavior of all workers is similar, clustering identifies two clusters, one with ....

[Article contains additional citation context not shown here]

Nickolayev, O. Y., Roth, P. C., and Reed, D. A. Real-time Statistical Clustering for Event Trace Reduction. International Journal of Supercomputer Applications and High Performance Computing (1997).

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