| | 1 | = [wiki:Testbed Testbed] = |
| | 2 | |
| | 3 | Our experimental platform included two identical hosts (one as primary and the other as backup), each with an IA32 architecture processor (Intel Core 2 Duo Processor E6320, 1.86 GHz), and 3 GB RAM. We set up a 1 Gbps network connection between the two hosts, which is specifically used for migration. In addition, we used a separate machine as a network client to transmit service requests and examine the results based on the responses. We built Xen 3.4 from source, and let all the protected VMs run guests OS (Linux 2.6.18 kernel). The file system of which is an image file of 3 GB shared by |
| | 4 | two machines using NFS. |
| | 5 | |
| | 6 | Our experiments used the following VM workloads under Primary-Backup model: |
| | 7 | |
| | 8 | Static web application: Apache 2.0.63. Both hosts are configured with 100 simultaneous connections, and repetitively downloaded a 256KB file from the web server. Thus, the network load will be high, but the system updates are not so significant. |
| | 9 | |
| | 10 | Dynamic web application: SPECweb2005 benchmark comprises of a web server serving a complex page request (e.g., CGI scripts), among other features. Both hosts generate a load of 100 simultaneous connections to the web server. |
| | 11 | |
| | 12 | Memory-intensive application: Since the FGBI is proposed to solve the long downtime problem under LLM system, especially when running heavy computational workloads on guest VM, we continued our evaluation by comparing FGBI with LLM/Remus under a set of industry-standard workloads, specifically SPECsys and NPB. |
| | 13 | |
| | 14 | To ensure that our experiments are statistically significant, each data point is averaged from twenty sample values. The standard deviation computed from the samples is less than 7.6% of the mean value. |
