[LINK] Knight of razor to slash government spending

Glen Turner gdt at gdt.id.au
Wed Apr 30 13:50:30 AEST 2008


Virgil wrote:
> Is the resource allocation capability of any virtualisation platform really the problem ?  
> Or inaccurate assignment of those resources by the people creating/administering the platform?

But the assignment is necessarily inaccurate, as the virtualisation environments
don't provide the necessary tools for an accurate assignment.

> And how is that any different from those same people not allocating the 6 spindle LUN to the
 > application when it's hosted on a physical server instead of a virtual server ?

The difference is in the extrapolation. I can buy a 6 spindle 1RU server,
4 spindle machines are the norm. But a 32 disk array is a serious bit
of kit, costing $30K.  And we want two of those to replace 10 x 1RU servers
(at $2600 each). As a result people underspend on the number of spindles in
virtualised systems.

Now I totally agree that not all of those 10 servers would be doing useful
stuff, but they contend for spindles with the servers which are. So if you're
going to underprovision you'd better have a really good handle on the
capacity planning needs of each server's applications.  Which, at the
moment, system administrators have no visibility of.

> The capabilities and performance in current enterprise virtualisation platforms allow running
 > (critical) production workloads today.

I don't see that. I see capacity planning failures rising from near-zero
to being probably the third major reason for system failures I've
experienced this year (after gross misconfiguration and security compromise).

Granted the places where I hold accounts are all universities, but
they're usually ahead of the curve rather than being unique.

 > But as you stated, if you want to realise more of the benefits of consolidation, you have
 > to engage in a realistic capacity planning exercise.

Sure, and my whinge was that whereas you can read capacity planning details
in mainframe documentation (where virtualisation has been used since the
1980s), there's nothing available to today's system administrators but
suck it and see.  Which is a tad incompatible with the notion of 'critical'.

Capacity planning isn't so much of a problem on real hardware, since there's
so much over-capacity. So it's become a forgotten art, and until vendors and
system administrators re-learn it and until tools with the same capabilities
as mainframe capacity planning tools appear then virtualisation of production
servers is problematic.

-- 
  Glen Turner



More information about the Link mailing list