FUENTE: AAS-X

The network speed specified the bandwidth and latency between individual client and servers and an interconnection2. In the experimental setup used here the interconnection was represented by a central router via which all IP traffic was routed. Figure 8.1 shows the logical network topology used to represent a simplified distributed storage system.

Figure 8.1: A simplified distributed storage system.

The network speed specified latency and bandwidth between client and servers and the cen- tral router. Bandwidth and latency applied to a specific link was the same in both directions. Two categories of network speed configurations were defined, a uniform and a tempo- rally varying one. In uniform network speed configurations, all servers exhibited the same network speed; they were used to simulate scenarios in which a network bottleneck was exhibited on the client or server side or on neither side.

Simulated scenarios with uniform network speed configurations:

• bottleneck on client side: The bandwidth between the client and the central router was significantly lower than between the individual servers and the central router. The latency between the client and the central router was significantly higher than between the individual servers and the central router.

• bottleneck on server side: The bandwidth between the client and the central router was significantly higher than between the individual servers and the central router. The latency between the client and the central router was significantly lower than between the individual servers and the central router.

• no bottleneck: All links between the individual participants and the central router exhibited the same bandwidth, no latency was configured.

Simulated scenario with temporally varying network speed configuration: The net- work speed varied between individual links to the central router and also varied over time. The configured bandwidth and latency values were pseudo-randomly selected from ranges of bandwidth and latency values, defined by maximum and minimum values. This pro- vided variation between the network speeds of the individual links during an experimental run, but allowed reproduction of the same network speed patters between repetitions of ex- perimental runs. A high bandwidth value directly correlated with a low latency value. For instance, a configuration with a maximum bandwidth of 100 Mbps, a minimum bandwidth of 1 Mbps, a maximum latency of 100 ms and a minimum latency of 1 ms would result in a latency of 1 ms if a bandwidth of 100 Mbps were selected. At regular intervals a new network speed configuration was generated on every link.

Implementation: The network traffic shaping was implemented on the central router fol- lowing guidelines from [38, 37, 35] using the Linux traffic control softwaretc, as available inCentOS 5(kernel 2.6.18-8.el5).

8.4

Experiment Setup

This section reports on the configuration values of the experimental parameters, including the churn pattern, workload, network speed, data item size and the autonomic manager configurations. Additional preliminary work is reported in appendix B.1, including the derivation of specific configurations and evaluations of the experiment harness.

8.4.1

The Test-Bed

The experiments were carried out in the same test-bed as used for P2P layer experiments, described in section 7.4.1. The network configuration was adapted so that one of the com- puters in the test-bed acted as the central router (see section 8.3.4). The client was config- ured withglobal knowledgeof all storage server addresses and their key ranges.

8.4.2

Derivation of User-Level Metrics

These experiments were carried out to evaluate the effect on performance and resource consumption of specific policies which varied the DOC in response to specific conditions. Performance was measured in terms of the time it took to completegetrequests3_{, the error} rate and the time after which failinggetrequests reported an error. The network usage was obtained from the Linux traffic shaping softwaretc.

Above performance measurements were combined to give an expected get timeusing the same principles and motivations as for computing an expected lookup timein the experi- ments with ASA’s P2P layer (see section 7.4.2, formula 7.1). To obtain the expected get time, theget time, theget error rateand theget error timewere aggregated over observa- tion periods of five minutes. The network usage measurement differed from earlier P2P layer-specific experiments due to a change in the version of theRAF DAmiddleware. A new version in which the network traffic monitor was not implemented anymore was used here. Thereforetc-statistics were used to record the total amount of data sent from storage

servers to the central router. Each experiment (specified by churn pattern, workload, data item size, network speed and policy) was repeated three times to verify the reproducibil- ity of the observed effects. As in the P2P layer experiments, all expectedget time values were aggregated over all three experiments and averaged in order to present a single per- formance measurement per experiment. The network usage was also specified by the mean of the three measurements (one measurement was available per experimental run).