On 3/12/2018 3:22 AM, Deepak Goel wrote:
A single OS and JVM does not scale linearly for higher loads. If you have
seperate OS and Java, the load is distributed across multiple instances
(with each instance only requiered to support a smaller load and hence
would scale nicely)
I had found this for running multiple apache servers on multiple VMs as
compared to a single instance (not Solr). But i am pretty sure it would be
same for Solr too
I think this is the last thing I'm going to say on the subject. You
disagree with a fundamental hardware concept that I've learned through
experience, so I might never convince you of anything. If that's the
case, I'm done trying after this, and I wish you the best of luck with
your efforts.
If the physical hosts you put the VMs on are far more powerful than you
would ever use for bare metal, and/or you split virtual machines between
different physical hosts, then that installation might scale better than
a single bare metal host. The decision makers in most companies are
usually a lot more willing to buy really expensive hardware if you tell
them it's for virtualization than they are for a single-purpose machine.
But if the bare metal environment has the same number of physical
servers with the same specifications, then a well-tuned bare metal setup
is going to perform better than virtual machines. There's nothing wrong
with VMs. They can perform very well if everything's sized appropriately.
Anytime a virtualized environment performs better than bare metal, it's
usually going to be that way because the virtualized environment has
different hardware than the bare metal environment. That hardware will
probably be much more expensive, and/or newer hardware that just works
better. It might also happen because the software installation was not
set up correctly to fully utilize all the hardware.
Solr works best with a lot more memory installed than people usually
install, *especially* with virtual machines, where RAM may be an even
more precious commodity than it is in bare metal servers.
Thanks,
Shawn
