Intel 7400 manual Scenarios 2 and 3 Virtualized with and without VMDq/NetQueue, vSW1

White Paper Consolidation of a Performance-Sensitive Application

Client 1	1 GbE			ESX* Server
					VM1
		Force 10 S50*	10 GbE
				vSW1
Client 8	1 GbE					VM8

Figure 11.. Virtualized lab test setup..

	0.30
(rms)	0.25
latency	0.20
Average	0.15
	0.10
	0.05

64		256		1024
		Packet size (bytes)
	Native		VMDq	No VMDq

Figure 12.. Netperf 2..4..4 UDP latency test with eight parallel streams..

Scenarios 2 and 3: Virtualized with and without VMDq/NetQueue

In these scenarios, the setup shown in Figure 11, consists of eight clients connected to eight 1-GbE ports of a 1G/10G link aggregation switch (Force 10 S50) and the Intel Xeon processor 7300 server connected to 10G port of the switch via Intel 82598 10GbE CX4 NIC. On the server, ESX 3.5 is installed and there are eight virtual machines created. The virtual machines are configured with 1vCPU; 1 GB RAM and SLES 10 SP1 is the guest operating system. VMDq with 16 queues along with NetQueue is enabled on the ESX server, and we are letting the VMM handle the VMDq assignments, core assignments, and interrupt affinity. On the clients, SLES 10 SP1 is the operating system. There are eight parallel streams of UDP latency tests being run from eight clients to eight virtual machines.

As mentioned earlier, in all the scenarios we used Netperf 2.4.4 UDP latency test and we ran the tests for UDP packet sizes of 64 bytes, 256 bytes, and 1024 bytes. The ESL workload used UDP packets of sizes varying from 40 to 200 bytes.

The results from these tests are summarized in Figure 12. In the graph, we are comparing native, virtualized with VMDq, and virtualized with no VMDq (represented by the blue, light blue, and gray bars, respectively). The horizontal axis represents various UDP packet sizes in bytes, and the vertical axis represents the average latency in milliseconds.

From Figure 12 it can be concluded that virtualization increases the latency. In fact, the latency is doubled (for 64-byte packets, the latency in the native scenario is 0.12 ms, whereas in the virtualized­ scenario with no VMDq it is 0.24 ms). By enabling VMDq and NetQueue, the latency in the virtualized case is near native (~0.13 ms). The increase in latency by virtualizing has a negligible impact when compared to in-game latency of 5 ms best case.

9