E. Maillet and C. Tron, "On efficiently implementing global time for performance evaluation on multiprocessor systems," Journal of Parallel and Distributed Computing, vol. 28, pp. 84-93, July 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....length of the message passing phase necessary to reduce the error will become prohibitively long. actual clock correction Figure 2: Estimation Error Grows as Experiment Time Grows The solution to this problem is to obtain data points for synchronization both before and after the experiment [22]. This will eliminate the possibility of performing on line adjustment of timestamps. However, that is not a concern in our case anyway since the global timeline is only used in postanalysis. Because sampling is done both before and after the experiments, the error in clock adjustment cannot grow ....

E. Maillet and C. Tron, "On efficiently implementing global time for performance evaluation on multiprocessor systems," Journal of Parallel and Distributed Computing, vol. 28, pp. 84-93, July 1995.

.... node (value returned bythephysical clockof the processor) POM also incorporates a mechanism for dating events globally.We opted for an approach based on a statistical method that consists in estimating the drift of the physical clock of each application node with respect to a reference clock [8]. Once the characteristics of the clock drifts have been determined for each application node, it is possible to relate the dates taken on the clocks of the application nodes to that of the reference node. The accuracy of the global time obtained this wayis 3 sufficient to ensure a coherent ....

E. Maillet and C. Tron. On Efficiently Implementing Global Time for Performance Evaluation on Multiprocessor Systems. Journal of Parallel and Distributed Computing, 1994.

....des differentes horloge locales des noeuds d application par rapport a une horloge de reference. L idee de base est celle exposee dans [HAD 88] elle a ete pour la premiere fois implantee sur machine parallele pour Echidna [JEZ 89] et une variante, developpee dans l equipe Apache a Grenoble [MAI 95] a ensuite ete integree a la POM. Une fois determinees les caracte ristiques de derive des horloges des differents noeuds d application, on est en mesure de ramener les dates prises sur les horloges des differents noeuds d application a celle du noeud de reference qui est alors la reference de ....

MAILLET E. et TRON C. On Efficiently Implementing Global Time for Performance Evaluation on Multiprocessor Systems. Journal of Parallel and Distributed Computing, 1995. A paratre.

....evidence [ 25 ] shows that crystal clocks drift linearly if the temperature regime is constant and that anyway drift is small ( 10 Gamma5 ) so that second order terms are neglected in the following. For runs longer than a few minutes temperature oscillations may occur for transputer clocks [ 26 ] in which case second order corrective terms offer one solution. It is therefore assumed that for correct clocks the accuracy of clocks should be such that (1 Gamma f) t Gamma t 0 ) Gamma ae H(t) Gamma H(t 0 ) 1 f) t Gamma t 0 ) ae; 1) where f is the maximum clock frequency ....

E. Maillet and C. Tron. On efficiently implementing global time for performance evaluation on multiprocessor systems. Journal of Parallel and Distributed Computing, 28(1):84--93, 1995.

....those tools have complete insight into the trace information of the application. The controlled logical clock requires clock differences to be limited to about 2 ms by a previous synchronization. This can be achieved by an explicit synchronization before and after the run of the application [20] and a linear interpolation as done by several monitor tools , or it can be done by a resources saving continuous synchronization, e.g. with xntp [22] The controlled logical clock does not modify the timestamps of the original clocks C i as long as the timestamp of a receive event is later ....

....is not finished until process B sends its message at t = 3. This example shows that the logical clock is not sufficient for visualizing monitor information. A representation without backward references can be achieved alternatively with a sufficiently exact synchronization of the local clocks [15, 18, 20]. Often used methods are an exact synchronization and drift estimation before the beginning of the application [8, 9] or better before and after the application [20] with a linear interpolation while the application is running. The exact synchronization can be done by deterministic [28] or ....

[Article contains additional citation context not shown here]

E. Maillet and C. Tron. On efficiently implementing global time for performance evaluation on multiprocessor systems. Journal of Parallel and Distributed Computing, 28:84--93, 1995.

....where workstations of different computing power work together. 4 The performance monitoring facilities in LOGFLOW There are three main problems that should be solved in any distributed monitoring environment: the intrusion problem (or probe effect) 15] the clock synchronization problem [14] . collection of enormous amount of information The intrusion problem means that the monitoring of nondeterministic programs influences their time dependent behaviour. Also the measured times of events are modified by the monitoring system and hence the performance values derived from these time ....

E.Maillet and C.Tron, On efficiently implementing global time for performance evaluation on multiprocessor systems, Journal of Parallel and Distributed Computing, 28(1) (1995) 84-93. 34

No context found.

E.Maillet, C. Tron. On efficiently implementing global time for performance evaluation on multiprocessor systems. J. of Parallel and Distributed Computing, 28, 84-93, 1995.

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