Intel 7400 Server hardware, Network I/O, Result NIC performance can be up to ~60% underutilized

White Paper Consolidation of a Performance-Sensitive Application

VM	1		VM	VM	n
		2
Virtual			Virtual	Virtual

Virtualization

Hypervisor

NIC

LAN

Throughput

10.0
8.0
	Unused
6.0	I/O capacity
4.0
	4.0
2.0

without VMDq

Result: NIC performance can be up to ~60% underutilized

Figure 3.. Network data flow for virtualization without the use of VMDq and NetQueue technologies..

Figure 4.. Impact of virtualization on a 10 GB Ethernet NIC without the use of VMDq and NetQueue..

of virtualization in the Intel lab using common network micro- benchmarks before attempting the virtualization of the gaming server environment. This would allow us to quantify the latency added by virtualization to see if it would be significant. When we were sure that the latency added should not be a concern, we proceeded to test the gaming server virtualization with private testing in the ESL lab and ultimately onto public testing on the Internet with real ESL members.

Server hardware

The PoC targeted the Intel Xeon processor 7300 platform with four processor sockets with the six-core Intel Xeon processor 7400 series (Dunnington). These new processors became available in September 2008 and are hardware and software compatible with Intel Xeon processor 7300-based platforms that have been in production for more than a year. The Intel Xeon processor 7400 series delivers a performance boost from using six rather than four cores per socket and by the addition of a new 16 MB L3 cache. It also delivers an energy-efficiency boost derived from our 45nm high-k process technology. In addition, the Intel Xeon processor 7400 series has added some

enhanced hardware-assist features for virtualization. The platform supports 32 memory slots for up to 256 GB capacity. In this PoC we used 32 GB.

Network I/O

But virtualization is not just about CPU and memory resources. It’s important to have I/O tuned for virtualization, too.

In a typical virtualization scenario (Figure 3), the network I/O for all the VMs is delivered to the hypervisor. The hypervisor then performs the necessary Ethernet switching functions in software to forward each network flow to the destination VM. This software function, called a virtual switch, is much slower than

atypical hardware-based Ethernet switch and causes CPU loading that detracts from application VM performance. Also, the hypervisor virtual switch has to process all the interrupts sent by the network I/O device on a single CPU core. This can be a bottleneck too, especially for faster networks like 10 GbE.

As shown in Figure 4, the Intel® 10 GbE NIC runs into this single- core interrupt processing load bottleneck. In this case, the 10 GbE NIC can only receive 4 GB of traffic due to the saturation of the single CPU core processing all the receive interrupts at 10 GB line rate.

5