[Beowulf] VMC - Virtual Machine Console

[Joe Landman]

Meng Kuan wrote:

> We performed some benchmark testing with linpack and bonnie++ on the
> VM and on the physical host. For para-virtualized VMs, the linpack
> performance is on par with the physical host. However, for bonnie++
> tests, para-virtualized VMs fell way behind physical host's
> performance. In short, CPU-bound and memory intensive HPC apps should
> do ok but not IO-intensive apps. More testing and fine-tuning will
> probably be needed to see how far we can push the VM in terms of
> IO-intensive operations but we are hoping that in time to come
> virtualization technologies will be able to narrow that gap.

Hi Meng:

   Not to ignite flammable substances here ... but there are a few 
hallmarks of HPC applications.  One of those is "beating the heck out of 
  a specific available resource".  Extra layers only add to this.

   What I want is thunking-free VMs.  It would be really nice to take an 
8 core workstation/server, run our base OS on one or two cores, and run 
other OSes on the other cores.  The problem is that this is not easy to 
do with todays commodity hardware.  Moreover, you pay a (sometimes huge) 
performance penalty for doing this, as you have single points of 
information flow (SPIF).  These SPIFs are anathema to HPC.  They are 
rate limiting.  They can increase contention/latency, decrease effective 
bandwidth.

  I like the idea of VMs for services that need HA, and for OSes like 
windows that need a safe place to run in.  HPC apps will stress one or 
the other portion of the machine.  They will beat on the memory 
bandwidth in some cases, which is why, despite AMD Opterons of old 
(single/dual core) having a disadvantage in computational performance to 
older Xeons of woodcrest derivation, they are still faster on specific 
memory bound problems and code (that second memory bus is hard to beat).

That said, and the point of this is that many HPC apps are rapidly 
becoming IO bound, as they need to move ginormous (meaning really large) 
amounts of data to and from disk, and MPI codes usually need to move 
data at the lowest latency possible.

There VMs which negatively impact IO performance (bandwidth/latency) 
will be problematic.

What would be interesting is a VM OS bypass for IO.  VM talk directly to 
hardware.   Not sure it is possible though, unless you are using a 
hypervisor, and a thin VM (OpenVZ?).

Just some thoughts, hopefully not all that flammable  (Jeff, what is 
that rule?  I am being asked, and I don't have an answer ...)


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
        http://jackrabbit.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 866 888 3112
cell : +1 734 612 4615