Hardware virtualization not all that…?

[Zaphod]

At least not yet, according to this paper from VMWare. An interresting (but slightly disappointing) read.

Intuitively, one would expect hardware support to improve performance across the board, but experiments on first generation hardware painted a mixed picture. Our software and hardware VMMs both perform well on compute-bound workloads. For workloads that perform I/O, create processes, or switch contexts rapidly, software outperforms hardware. In two workloads rich in system calls, the hardware VMM prevails.

We have traced this surprising result to its root cause by studying architecture-level operations for which the software and hardware VMMs impose very different amounts of overhead. While the new hardware removes the need for BT and simplifies VMM design, in our experiments it rarely improves performance. New MMU algorithms in the hardware VMM might narrow this gap, but current hardware support is CISC-like in its packaging of a whole solution and difficult to exploit without giving up existing software techniques.

 

[chavvdarrr]

tests were made using iP4 ... and according to

7. Software and hardware opportunities
......
7.4 Hardware MMUsupport
The picoJava microprocessor lands at one end of the hard-
We are optimisticabout the potential of future hardware assistance
ware/software split [15]. This processor effectively reimplements
in the area of MMU virtualization. The three-way trade-off among
the software JVM version 1.0 in hardware. While picoJava was
trace costs, hidden page faults and context-switch costs was suc-
neverintendedasahighperformancebytecode executionengine,it
cessfully addressed with appropriate hardware in IBM’s s/370 ma-
isstillstrikingtoobservejusthowdifferentlyamodernJVMdeliv-
chines [17]. Both AMD’s “nested paging” [2] and Intel’s “EPT”
ersperformancethroughmulti-stageJITcompilation.ThepicoJava
[16] proposals for MMU virtualization are similar to the SIE de-
design, frozen in time in the last millenium, is not what one would
sign.

we may soon see improvements - afaik Pacifica is coming with new Opterons and s.F

 

[Zaphod]

we may soon see improvements - afaik Pacifica is coming with new Opterons and s.F

Yeah. Thus the 'at least not yet' and only 'slight' disappointment. The final paragraph of the conclusion is somewhat more optimistic than the first two I quoted. We’ll see.

VMMs. We hope that coming hardware support for virtualization will be designed to blend easily with and complement existing software techniques. This may contribute to faster adoption of the hardware and advance the general state of virtualization.