J. K. Hollingsworth. Critical path profiling of message passing and shared-memory programs. IEEE Transactions on Parallel and Distributed Systems, 9(10), October 1998.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....in a parallel system is critical path analysis. By discovering the chain of dependent events that determined the overall execution time, critical path analysis has been used successfully for identifying performance bottlenecks in large scale parallel systems, such as communication networks [3, 9]. Out of order superscalar processors are fine grain parallel systems: their instructions are fetched, re ordered, executed, and committed in parallel. We argue that the level of their parallelism and sophistication has grown enough to justify the use of critical path analysis of their ....

J. K. Hollingsworth. Critical path profiling of message passing and shared-memory programs. IEEE Transactions on Parallel and Distributed Systems, 9(10), October 1998.

....in Fig. 1 spends most of the time in function b( and suggest that function for optimization. This is shown in Fig. 2 where the functions a( to d( are found at the top; the other functions shown are only for internal use in the Solaris operating system. The tools and methods described in [3, 8, 18] have critical path analysis and will show that a( is the most beneficial to optimize. However, those tools and methods assume that there is only one thread (or process) scheduled on each processor. Our approach with the extended critical path generalizes the two extreme cases above. Not only ....

....the Extended Critical Path with One Thread per Processor We start to look at the extended critical path algorithm under the condition that the program is executed on a sufficient number of processors and thus no threads have to share a processor at any time. This is the same assumption used in [3, 8, 18]. We assume that explicitly synchronized events happen at the exact same time, such as one thread releasing another thread. This means that a thread starts to execute at exactly the same moment as the other thread releases it. This assumption is realistic, does not change the algorithm in ....

[Article contains additional citation context not shown here]

J. Hollingsworth, Critical path profiling of message passing and shared-memory programs, IEEE Trans. Parallel Distrib. Systems 9, 10 (October 1998), 1029#1040.

....yet been developed. Our instrumentation scheme also gives the length of each subpath in a critical path. As far as we know, previous schemes either provide the length of only the critical path [1, 6] or provide a list of procedures and the amount of time each contributes to the critical path [3]. The usefulness of our scheme has been demonstrated through realistic applications running on symmetric multiprocessors. Instrumentation code has usually been inserted into the synchronization parts and entry exit points of procedures by programmers, but our scheme is implemented by the ....

....on a fully strict computation, it deals only with implicit synchronization associated with forkjoin primitives. It should also be noted that in the Cilk scheme, the shape of a DAG does not vary in different runs and the spawn and communication edges have zero weight. Paradyn. Hollingsworth s work [3] computes at runtime a critical path profile that gives us a list of procedures whose contributions to the critical path length are large. The language model Hollingsworth uses is essentially the same as ours, although the communication in his target language is performed not through channels ....

J. K. Hollingsworth. Critical Path Profiling of Message Passing and Sharedmemory Programs. IEEE Transactions on Parallel and Distributed Systems, pages 1029--1040, 1998.

....with a simple contention model. However, this simplistic model is inadequate to describe the performance of many parallel applications because of complex interactions between constituent processes. For example, we might use the critical path notion to take inter process dependencies into account [11]. Other application models could model piece wise linear curves. Figure 4 shows an example of Harmony s configuration choices in the presence of our client server database and an application with variable parallelism. The parallel application s speedup curve is described in an application specific ....

J. K. Hollingsworth, "Critical Path Profiling of Message Passing and Shared-memory Programs," IEEE Transactions on Parallel and Distributed Computing, 9(10), 1998.

....API. In addition to the ones described below, a few other projects around the world are using the API, but are at too early of a stage to report here. At the University of Maryland, we have used the API to implement an online version of critical path analysis for parallel programs running on SMPs[6]. In this application, instrumentation code to track the synchronization events in a shared memory program is inserted on demand using the API. At Maryland, we are also investigating using Dyninst as part of the Harmony project[7] The Harmony project is looking into using runtime observations of ....

J. K. Hollingsworth, "Critical Path Profiling of Message Passing and Shared-memory Programs," IEEE Transactions on Parallel and Distributed Computing, 9(10), 1998, pp. 1029-1040.

No context found.

J. K. Hollingsworth. Critical path profiling of message passing and shared-memory programs. IEEE Transactions on Parallel and Distributed Computing, 9(10):1029--1040.

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