You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 10 Next »

This is the start of the FY2006 DoE Terapaths DWMI Progress Report due September 10, 2006

Current draft

Template for Report

Current Text of Report

Terapaths: A QoS Collaborative Data Sharing Infrastructure for Petascale Computing Research: DWMI: Datagrid Wide Area Monitoring Infrastructure
Les Cottrell, Yee-Ting Li & Connie Logg, Stanford Linear Accelerator Laboratory (SLAC)

Summary:
The main goal of the DWMI project is to build, deploy and effectively learn how to use an initially relatively small but rich, robust, sustainable, manageable network monitoring infrastructure focused on the needs of critical HEP experiments such as Atlas, BaBar, CMS, CDF and D0.

Topics still needed:
More on Terapaths interactions.
More on why or how users/physicists care, even anecdotes, testimonials

We have room for a figure, it needs to be compelling.
Today's data intensive sciences, such as High Energy Physics (HEP), need to share large amounts of data at high speeds. This in turn requires high-performance, reliable end-to-end network paths between the major collaborating sites. In addition network administrators need alerts when there are anomalous events, and grid middleware and end-users need long and short-term forecasting for application and network performance for planning, setting expectations and trouble-shooting. To enable this requires a network monitoring infrastructure between the major sites that can help notify and identify potential problems.

Active monitoring: We have developed an active network monitoring toolkit (IEPM-BW). It provides measurements, data arching, analysis, reporting and visualization. This is now being used to make regular measurements from the following major LHC related sites: CERN, BNL, Caltech, FNAL, SLAC, and Taiwan. We also have about 60 locations worldwide that are being monitored from these important sites. We use a selection of probes based on the quality and interest in the path being measured utilizing metrics such as network routes, round trip time, one-way delays, available bandwidth and achievable throughput. We are extending the presentation of IEPM-BW by working with the USATLAS and ULTRALIGHT groups to customize reports on their most relevant interests.
To aid the detection of incorrect host settings, we have also implemented features to query the configurations of the network monitoring hosts.
As part of this we have developed and put into production, management tools for automation and robustness, including: installation and update kits; measurement and reporting of unreachable participating hosts; documentation, including a Program Logic Manual; and a database of site, host, location, contact, OS, cpu, test parameters.

Passive Monitoring: We have studied and reported on limitations using current active end-to-end network measurement techniques in future high-speed networks. As a result of this we are exploring the effectiveness of using passive (e.g. Netflow) tools to augment or even replace some of the active measurements. In conjunction with BNL we are building a netflow monitoring toolkit using open source software to bring together quality tools to gather, store, process, analyze and visualize the performance information. The intent is to make this generally available and deploy at LHC sites such as BNL, CERN, SLAC and Michigan.
In fact, much of our development is steared by the requirements of the BNL site where we have a development version of the entire suite running.
Event Detection and Diagnosis: It is increasingly impossible for network managers to manually review thousands of reports each morning to detect problems. Thus we are developing tools to automate this activity. This includes forecasting and comparing the observed with the forecast to detect anomalous events, reporting the events, and automatically gathering information relevant to the event to assist in diagnosis. One version of the event detection is in production use. As part of this, in the last year, we have also detected, reported (together with in-depth case studies) and helped diagnose major problems at sites such as BNL, Taiwan, SDSC, NRL, BINP, and CERN.

Do we want to mention that the code is still in development and that we are basically field testing the alerts to known problems and using as a basis towards event diagnosis?

High speed data transport: Our involvement in high speed TCP transport, including several publications and winning the Supercomputing Bandwidth challenges three times in succession and the Internet land-speed record twice has led Microsoft to request our help in evaluating their next generation TCP stack (CTCP). Given the extent of Windows deployment it is critical to ensure that CTCP performs well without a negative impact upon the iIternet community.
As part of this we have identified and aided the testing of numerous added features to aid the performance of the delay-based congestion control algorithm used in CTCP. Having finalized our initial report into the deployment impact of using CTCP in production environments on both long and short distance high speed Internet paths, we are know looking to publish a joint conference paper with Microsoft of our work.

Internet Measurement Confederation: An important aspect of being able to both understand and diagnose network performance problems is the unification of reporting formats and the understanding of tool performance on the Internet.
SLAC has recently started close collaboration with both Internet2 and ESnet to help develop and expand the functionalities of the international PerfSONAR collaboration.
Having had gained much momentum over the last few months due to its open-source, open-community, open-standards based ethos of network monitoring, SLAC were delighted to help contribute our network analysis skills and experience to apply the PerfSONAR technology to production systems like that of the LHC project.
We aim to apply much of our existing analysis frameworks and tools to benefit the PerfSONAR project - including that of event detection, event diagnosis.

===The rest of this is from Last Year's ===
For detecting drops in network performance (event detection) we have developed, published and integrated an automated step change detection algorithm. It has been successfully applied to several metrics with different measurement repetition frequencies, including RTT, available bandwidth and achievable throughput. It is now in regular use to generate email alerts for network administrators. With this, we are now reliably (few false positive or misses) detecting events within a few hours of their onset. This is in contrast to before, where in significant events went undetected for weeks.
Given the success and experience of these alerts, we are working on developing tools to gather more information to report to the network administrator. For the future we are evaluating other event detectors including the use of neural networks and Principal Component Analysis (PCA) to enable simultaneously evaluating multiple metrics and paths.
We have developed and are now integrating a long-term forecasting technique that takes into account seasonal (e.g. diurnal and weekly) variations. As part of the integration we will also make the forecasting tool more general purpose so it can be applied against data from other monitoring infrastructures.
In preparation for evaluating QoS at BNL we worked with ESnet to evaluate the impact and use of the ESnet OSCARS project. Our next steps will be to set up the measurements for the QoS project at BNL.
Impact to specific DoE Science applications
Improved network understanding and clearer expectations together with more quickly discovering and reporting network problems is critical to all network based applications. The DWMI project's deployment of the IEPM-BW infrastructure focuses on the real needs of the DoE supported LHC, BaBar, CDF and D0 HEP experiments and provides an evolving and practical basis for improved network monitoring and management.

Figure 1: Forecasting of available bandwidth and detection of a drop in performance event. The sudden drop in performance (below the expected range - black line) is automatically detected.

For further information on this subject contact:
Dr. Thomas Ndousse,
Mathematical, Information, and Computational Sciences Division
Office of Advanced Scientific Computing Research
Phone: 301-903-9960
tndousse@er.doe.gov

  • No labels