MPI collective operations are implemented using a variety of algorithms which define different communication patterns between the ranks involved in the operation. The performance of these algorithms in multi-core clusters highly depends on the mapping of the ranks to the system processors due to the uneven capabilities of shared memory and network channels. The hierarchical design of these algorithms contributes to use optimally the communication channels. Nevertheless, common hierarchical algorithms have shown themselves, for some collectives as allgather, inefficient and even impracticable. This paper analyzed the reasons for that and works out an alternate approach through performance modeling. Such approach, departing from the a priori knowledge of a regular mapping as round-robin or sequential, and keeping the original algorithm unmodified, switches at run time the rank to process mapping into another regular mapping that reduces network traffic. The methodology is evaluated with three collectives and their underlying algorithms, showing speedups of up to 5x in the Binomial Tree or 3x in Ring algorithms compared to unfavorable mappings.

Heterogeneous systems composed of CPUs and accelerators sharing communication channels of different performance are getting mainstream in HPC but, at the same time, they show a complexity that makes it difficult to optimize the deployment of a data parallel application. Recent analytical tools such as Functional Performance Models, combined with advanced partitioning algorithms, manage to achieve a balanced configuration by distributing the workload unevenly, according to the performance of the different processing units. Unfortunately, such uneven distribution of the computation load leads to communication unbalances that, very often, render worthless the previous workload balancing efforts. Finding the optimal communication scheme without expensive testing on the executing platform requires an analytical approach to the estimation of the communication cost of different configurations of the application. With this goal in mind, we propose and discuss an extension of the t-Lop communication performance model to cover heterogeneous architectures. In order to provide a quantitative assessment of this extended model, we conduct experiments with two representative computational kernels, the SUMMA algorithm and the 2D wave equation solver. The t-Lop predictions are compared against the HLogGP model and the observed costs for a variety of configurations, hardware resources and problem sizes.

Inter- and intra-machine MPI collective operations in current multicore clusters are essentially different, and therefore their performance modelling ask for different approaches. Inside a multicore each individual message transmission in a collective operation flow in parallel with others, but sharing the bandwidth of the main memory channel. Current models ignore this issue, making errors like giving the same cost estimation to quite different collective algorithms. We outline a new cost model focused on shared channels.

Current voice recognition systems tend to be implemented as embedded proprietary solutions. This model is not suitable for the growing complexities of present and future developments: It is single-user, it is non portable, and it assumes the workstation model, where all the CPU resources are supposed to he locally available. This work researches how a high performance speech recognition system can be redesigned and implemented as a time-critical network service with three main design goals: Scalability, predictability and POSIX portability. The whole idea has been tested by rebuilding IVORY, a well known robust desktop voice recognition methodology, as a distributed service. Also, potential fields of application are identified.

Abstract Current voice recognition systems tend to be implemented asa PC desktop facility. This model is not suitable for thegrowing complexities of present and future developments: Itis single-user, it is non portable, and it assumes theworkstation model, where all the CPU resources are supposedto be locally available. This work researches how a highperformance speech recognition system can be redesigned andimplemented as a time-critical network service shared throughordinary data transmission media with three main designgoals: Scalability, predictability and POSIX portability. Thewhole idea has been tested by rebuilding IVORY, a wellknown robust desktop voice recognition methodology, as adistributed component. 1. Introduction While Speech Processing and Recognition is a fieldexperiencing a rapid and promising expansion, the operating-system environments for the desktop PC still typically lack oftrue real-time support. To overcome this limitation, currentspeech recognition systems are confident on the workstationprinciple: all the CPU resources are always available to theapplication where they are embedded. This approach shows amain limitation: Its growing computational complexity.IVORY ([1], [7]), a stand-alone speech recognition system ofisolated words, gives figures of computational complexityaround 21 Mflop/s. Though this load is easily assumed bycurrent CPU's, continuous speech can raise the computingpower demand one order of magnitude. Noise cancellationdemands up to five or six times the power of the recognitionitself. Furthermore, new applications of speech processingdemand much more computing power. For instance, tracking asingle speaker by the Microphone Arrays technique shows acomputational complexity near 166 Mflop/s ([6]). Thoughtoday's PC microprocessors claim peak execution ratesexceeding 1 Gflop/s, regular DSP algorithms rarely result insuch a high performance. In our view, desktop speechprocessing is -and will always be- strongly limited by itscomputational complexity, nowadays constrained to thecomputing power of the average personal computer. This work was founded by CICYT and Junta de Extremaduraunder the TIC99-0609 (DIARCA) and CICYTEX IPRR98A039 projects respectively.Distributed computing should change this scenery.Ongoing developments on component based softwareengineering makes possible to envision a remote service ofDSP computing power for speech processing. It would allowto bring both to the current desktop PC and to the futureinternet appliances the more advanced developments on thefield. This work investigates the distribution of speechrecognition in the context of DIARCA, a research projectwhose aim is two-fold. Firstly, to distribute IVORY with threedesign goals: Scalability, predictability and POSIX portability.Secondly, to extend the results in order to support microphonearray developments. This work is about the first goal.Figure 1. IVORY: a Robust Voice Recognition system