Probes Coordination Protocol for Network Performance Measurement in GRID Computing Environment

The fast expansion of Grid technologies emphasizes the importance of network performance measurement. Some network measurement methods, like TCP throughput or latency evaluation, are very sensitive to concurrent measurements that may devalue the results. This paper presents the Probes Coordination Protocol (PCP) which can be used to schedule different network monitoring tasks. In addition, this paper goes on to discuss the main properties of the protocol; these being, flexibility, efficiency, robustness, scalability and security. This study presents the results of its evaluation and of experiment periodicity measurements. Introduction. The purpose of Computational Grids is to aggregate large collections of shared resources (computing, communication, storage, information) in order to build an effective and high performance computing environment for data-intensive or computing-intensive applications. The underlying communication infrastructure of these large scale distributed environments consists of a complex interconnection of the public Internet, local area networks and high performance system area networks like Myrinet. Consequently “the network cloud” may exhibit extreme heterogeneity in performance and reliability that can considerably effect the distributed application performance. In a Grid environment, monitoring the network is, therefore, critical in determining the source of performance problems or in fine tuning the system to perform better. For such purposes a network performance measurement system may be deployed over the Grid and net cost function may be computed and provided to a grid resource allocation component. Sensors that aim to measure the different network metrics such as end to end throughput, loss rate or end to end delay are the basic building blocks of the network performance measurement system. However, classical Internet measurement tools can also be used for this purpose. Two kinds of measurement methodology are classically applied (these being active and passive methods). Active methods inject extra traffic to determine the capacity of the links in terms of latency, loss or bandwidth. Passive methods measure the traffic but are unable to evaluate the real capacity of a link. For example to evaluate the available TCP or UDP throughput tools like Iperf [1] or Netperf [2] send probe packets during a given duration (default 10s). Amounts of send probe data are from 12.5 MB on a 10Mbps link to 1.25 GB on the 1Gbps link. As the traffic generated by active testing is added to the usual traffic load on the network, there are drawbacks to active methodologies. Firstly, they add a potentially burdensome load to the network; secondly, the additional traffic may perturb the network and devalue the resulting analysis. Therefore, these tools must be appropriately scheduled to minimize the impact on networks while still providing an accurate measurement of a particular network metric. Grid network monitoring raises problems which are not so critical in classical Internet performance measurement. As the number of sites and their respective logical links used by a community of users in a Grid environment are finite the probability of concurrent measurements is high. The possibility of sensor probes colliding and thereby measuring the effect of sensor traffic increases quadratically with the number of sensors [3]. This can become highly critical in hierarchical Grids like the HEP physics DataGrid (see section 3), organized following a multi-tiered architecture. For example, the probes from tier 1 to tier 0 (CERN) may collide frequently making tier 0 site a real bottleneck (see fig. 0.1). This leads to chaotic results. In this paper the main features of the protocol which we have developed for coordinating the network monitoring probes in the European Data Grid (EDG) project [5], are described. The paper is organized as follows. In the first section probes coordination service and four possible experiment scheduling strategies as well, as requirements which must be fulfilled for probes coordination service, are discussed. The second section goes on to describe probes coordination protocol, its design principle and characteristics. In a third section our implementation of PCP, with EDG distribution, the evaluation methodology and the results of testing are presented. Finally related work is discussed. CNRS-UREC, ENS-LIP, 46, allée d’Italie, 69 364 Lyon, France ([email protected]) INRIA-RESO, ENS, 46, allée d’Italie, 69 364 Lyon, France ([email protected]) CNRS-UREC, ENS-LIP, 46, allée d’Italie, 69 364 Lyon, France ([email protected]) INRIA-RESO, ENS, 46, allée d’Italie, 69 364 Lyon, France ([email protected]) 71 72 R. Harakaly, P. Primet, F. Bonnassieux, B. Gaidioz