| 18 | The first session (which occurred last) began by looking at the problem of |
| 19 | monitoring power usage information for testbeds networks and clouds. The |
| 20 | session began with a description of the Cloudlab power monitoring system |
| 21 | which uses the Intelligent Platform Management Interface (IPMI) to collect |
| 22 | power measurements over time and make them available to experimenters via a |
| 23 | web page. A related talk described the power monitoring service supported by |
| 24 | the Kwapi system which was designed to offer high frequency data collection, |
| 25 | allowing users to correlate and associate power measurements with other |
| 26 | network behaviors (e.g., bandwidth usage). |
| 27 | The federation of testbeds presents additional monitoring challenges |
| 28 | ranging from the complexities of dealing with heterogeneous infrastructure, |
| 29 | software systems, and data formats, to scalable data collection and |
| 30 | aggregation, to trust/reliability issues. Example systems such as |
| 31 | MODACLOUDS, SPECS, and DICE were described that use monitoring architectures |
| 32 | consisting of many distinct components, each with a specific purpose, to |
| 33 | address the complexity of these systems. The FELIX monitoring service |
| 34 | handles the scalability challenge by arranging measurement services |
| 35 | hierarchically at the highest level, but then using a distributed approach |
| 36 | within testbed islands. The Federated Trust and User Experience framework |
| 37 | address the problem of assessing a testbed's reliability and trustworthiness |
| 38 | by combining conventional monitoring information (objective information) with |
| 39 | reputation-based trust management systems (subjective information) to |
| 40 | define a reputation value that can be used to compare testbeds. |
| 41 | While much work has focused on monitoring testbed infrastructure, |
| 42 | relatively few tools are available to users to measure their |
| 43 | specific experiment or slice. Looking at experimenter mailing lists and |
| 44 | discussion groups shows that the majority of problems that experimenters |
| 45 | face would not be detected by monitoring the infrastructure, but rather |
| 46 | require slice/experiment-specific measurement and validation tools designed to |
| 47 | test and evaluate the performance of a user's slice or experiment. |
| 48 | |