Virtualizing I/O Devices on VMWare Workstation's Hosted Virtual
Machine Monitor - Sugerman et al - 2001

Kind of Paper:
--------------
   -- new approach (to virtualizing PC platform and I/O devices)
   -- design, implementation, and performance analysis

General Summary:
----------------
This paper describes VMWare Workstation's approach to virtualizing PC hardware
and more specifically the I/O devices. Since PC usually come with a 
preinstalled OS the authors suggest using it to host a user-level application
(VMApp) and a privileged virtual machine monitor (VMM). The basic idea is that
an arbitrary off-the-shelf OS (guest OS) can be put on top of the interface
exported by VMApp and VMM. VMM provides a efficient CPU virtualization while
VMapp uses the host OS services to virtualize I/O devices. In order to
demonstrate some of the problems that arise in virtualizing I/O devices, the
authors describe in detail a virtual Ethernet adapter and evaluate its
performance. The experimental results indicate that reasonable performance is 
achievable but only after some optimizations. 

Details:
--------
   -- Hosted virtual machine architecture
      == host OS - normal OS running on bare hardware
      == VMM - runs directly on hardware and virtualizes the CPU resources
      == VMApp - virtual machine application on top of host OS
   
   -- virtualizing I/O devices  
      == all I/O requests of the guest OS are intercepted by VMApp and VMM
      == I/O is handled by VMDrivers loaded into the host OS
      == the switch between VMM and the host OS is expensive
   
   -- optimizations 
      == eliminating unnecessary switches by emulating the OUT instruction in a
      way that requires a switch only when real I/O is required 
      == send packet combining
      == augmenting the VMNet driver with shared memory
      == modifying guest OS - improving context switch operation
      == optimizing guest driver protocol - idealized virtual drivers
      == modifying host OS - to avoid double buffering
      == bypassing host OS - special drivers that talk directly to the
      devices in VMM
      
Comments on Ideas and Methodology:
----------------------------------
   -- Virtualizing system resources indeed does provide lots of flexibility
   (different OS on different VM) and fault isolation. I would have like to see
   a bit more motivation though. 

   -- Using the host OS to do most of the "hard" work for I/O operations is a
   good idea. The paper makes the interaction between MVApp, VMM, and host OS
   clear only in the case of I/O. It is not quite clear how other resources
   (memory for example) will be jointly managed especially if the host
   operating system is allowed to run other applications besides VMApp.
   
   -- The network card virtualization, chosen for this paper, is a good way of
   illustrating some of the problems associated with I/O device virtualization.
   The reason is that network devices are expected to provide high throughput
   and low latency. In a way, this is a worst case analysis.
   
   -- The performance evaluation section is very good. All results are broken
   down as much as possible so that it is quite easy to identify the
   bottlenecks. The evaluation of the optimizations is done very systematically
   and the effects of each optimization are shown (table 3). Also, the 
   measurement procedures are described in lots of details which is important
   for reproducibility of results.

   -- The section describing the different optimizations was very well
   organized. It not only presented the benefits but also the disadvantages of
   each optimization. 
   
   -- The performance section could have offered a comparison with an 
   alternative architecture in which all I/O is handled by the VM monitor. 
   This would have helped in assessing the I/O device virtualization method
   chosen by the authors.