Benchmarks: ESX Server 3.0.1 vs XenEnterprise 3.2 networking performances comparison

VMware is decided to play the benchmarks game up to the end, and seems to prefer XenSource as sparring partner.

After publishing a much controversial performance comparison between ESX Server 3.0.1 and Xen 3.0.3 (which received a remarkable counter-analysis by XenSource, comparing ESX Server 3.0.1 against XenEnteprise 3.2 beta), now Palo Alto company is back and publishes another study focused on networking performances this time, against XenEnterprise 3.2:

Just a few years ago, generating enough network traffic to push a single 1 Gbps network adapter to capacity would saturate a computer’s main processor. However, processors have advanced in these few years to the point that 1 Gbps networking does not stress a modern native or virtualized system.

In order to better expose the real virtualization overhead of high-throughput networking, it is now necessary to increase the load on the system beyond a single 1 Gbps link.

In this paper, we characterize the virtualization overheads of virtual machines by measuring the performance under heavy networking in a uniprocessor virtual machine. This was done by configuring multiple 1 Gbps Ethernet adapters (“NICs”) in the server, each associated with a netperf or netserver instance running over a unique subnet and port.

The hypervisors tested were ESX Server 3.0.1 (referred to as “ESX301”) and XenEnterprise 3.2.0 (referred to as “XE320”). Both hypervisors were installed with no modifications or tuning.

For each hypervisor, the corresponding “tools” packages containing paravirtualized (PV) network drivers was installed in the guest. The hypervisors were installed on a 4-core, 3 GHz HP DL380G5 system with 16GB of memory. For the virtualized tests, a single virtual machine was configured for each hypervisor with one processor and 1GB of memory, running Windows Server 2003 Release 2 Enterprise Edition (32-bit). During the tests, the virtual machine under test was the only one running in the hypervisor…

Read the whole paper at source.