Evaluation

This section presents a quantitative evaluation of the Denali isolation kernel. We ran microbenchmarks to (1) quantify the performance of Denali's primitive operations, (2) validate our claim that our virtual architecture modifications result in enhanced scale, performance, and simplicity, and (3) characterize how our system performs at scale, and why. As previously mentioned, none of these experiments were run with the virtual MMU enabled. Additionally, in all experiments, our VMs ran with their data in virtual core, and as such, they did not exercise the virtual disks.³

In our experiments, Denali ran on a 1700MHz Pentium 4 with 256KB of L2 cache, 1GB of RAM, an Intel PRO/1000 PCI gigabit Ethernet card connected to an Intel 470T Ethernet switch, and three 80 GB 7200 RPM Maxtor DiamondMax Plus IDE drives with 2 MB of buffering each. For any experiment involving the network, we used a 1500 byte MTU. To generate workloads for network benchmarks, we used a mixture of 1700MHz Pentium 4 and 930MHz Pentium III machines.