Scalable dynamic monitoring

Scalable Dynamic Monitoring, Analysis and Tuning Environment for Parallel Applications66
P. Caymes-Scutari∗,a , A. Morajkob , T. Margalefb , E. Luqueb
of Information Systems Engineering, Universidad Tecnol´gica Nacional o Facultad Regional Mendoza. Rodr´ ıguez 273 (M5502AJE) Mendoza, Argentina. b Computer Architecture and Operating Systems Department. Universitat Aut`noma de o Barcelona. EdificiQ - Campus Bellaterra (08193) Barcelona, SPAIN.
a Department

Abstract Parallel/distributed systems are continuously growing. This allows and enables the scalability of the applications, either by considering bigger problems in the same period of time or by solving the problem in a shorter time. In consequence, the methodologies, approaches and tools related to parallel paradigm should bebrought up-to-date to support the increasing requirements of the applications and the users. MATE (Monitoring Analysis and Tuning Environment), provides automatic and dynamic tuning for parallel/distributed applications. The tuning decisions are made according to performance models, which provide a fast means to decide what to improve in the execution. However, MATE presents some bottlenecks as theapplication grows, due to the analysis process is made in a full centralized manner. In this work, we propose a new approach to make MATE scalable. In addition, we present the experimental results and the analysis to validate the proposed approach against the original one. Key words: dynamic tuning, performance evaluation, scalability, parallel applications

1. Introduction Every day, weexperiment the increase in the demand for high performance computing. Different areas of the science and the engineering tackle very complex problems, both from the data and the operation points of view. Parallel/distributed computing constitutes a very useful solution to provide the computational power required to solve the problems. As more and more organizations have high-speed local area networksinterconnecting many generalwork has been supported by the MCyT under contract TIN2007-64974. author. Phone: +54 261 5244 554, Fax: +54 261 5244538 Email addresses: pcaymesscutari@frm.utn.edu.ar (P. Caymes-Scutari), anna.morajko@uab.es (A. Morajko), tomas.margalef@uab.es (T. Margalef), emilio.luque@uab.es (E. Luque)
∗ Corresponding 6 This

Preprint submitted to Journal of Parallel and DistributedComputing

March 26, 2009

purpose workstations, the combined computational resources may exceed the power of a single high-performance computer. The most important factor in distributed computing is cost. The users see very little cost in running their problems on a local set of existing computers [7] even though the set is heterogeneous and/or non-dedicated. Nonetheless, developingparallel/distributed applications constitutes a challenging task due to it involves not only the parallel programming models and the communication libraries but it also includes some additional aspects such as decomposition, mapping, concurrency, scalability, efficiency, synchronization, etc., which determine the correct behaviour of the application [4]. Certainly, programming an efficient application to reallytake profit from parallelism, is an effortful challenge which requires a high degree of expertise. Considering these difficulties, users in general have to face the optimization or tuning process with the aim of improving the application and its behaviour. Happily, along the years different approaches were proposed and many tools were developed to assist the users in some phase of the optimizationprocess: monitoring, performance analysis or tuning. One of the performance approaches is the dynamic tuning approach, which is specially suitable for heterogeneous or time-sharing execution environments, due to monitoring, analysis and tuning processes are executed on the fly, with the aim of adapting the behaviour of the application to the current conditions of the system. There exist different...