Reducing Network Virtualization Overheads

Table 2: Distribution of CPU time during network transmission, with VMM optimizations. Many of the VMM time entries now represent a collection of individual instructions, which renders the Average Time not applicable.

Categories marked with `

' are partly derived from direct measurements presented in Section 3.5 for reasons described below.

Total Time
Category	Percent Time	Average Time
VMM Time	71.5%	N/A
Transmitting via the VMNet	17.9%	22.7 $\mu$ s
Receiving via the VMNet	2.5%	22.7 $\mu$ s
Emulating the Lance transmit path	1.8%	3.0 $\mu$ s
VMM Time
Category	Percent Time	Average Time
Guest idle	30.4%	N/A
Instructions not requiring virtualization	22.2%	N/A
Guest context switches	11.5%	N/A
Host IRQ processing while guest idle	10.7%	N/A
Virtualizing privileged instructions	7.9%	N/A
IN/OUTs to the PIC (Interrupt Controller)	2.5%	0.78 $\mu$ s
Virtualization overheads of guest IRQs	2.5%	N/A
IN/OUTs to the Lance status register	2.3%	0.91 $\mu$ s
Transitioning to/from virtualization code	1.5%	N/A
IN/OUTs to the Lance that world switch	0.5%	N/A

**Figure 6:** Throughput vs. Data Size when transmitting. The native and VM/**PC-733** lines are transmitting to **PC-350** and the VM/**PC-350** lines are transmitting to **PC-733**. The optimized VM/**PC-733** was able to achieve native speeds, but non-I/O virtualization overheads limited VM/**PC-350**'s achievable I/O throughput.
$\begin{figure} \centerline{\psfig{figure=plots/send.ps,width=6.7in,angle=270}} \end{figure}$