[COLUG] MVMRUG Spring Meeting Announcement - April 25 - Columbus OH

[Rick Troth]

On Sun, 20 Apr 2008, Travis Sidelinger wrote:
> I notice the large difference in user time, but linux on Z completed in
> less real time.  Any ideas why that would be?

I presume that Russ's system is running in a virtual machine.
The user and system numbers for 'time' in that context are
easily skewed except for the most recent kernels.  Consider this:
the hypervisor dispatches our "guest", guest kernel kicks into user
mode and notes that time, hypervisor dispatches a different guest
or two (or twenty), hypervisor then re-dispatches our guest of interest,
which finishes the user timeslice and notes that time.  The actual time
of the slice is much less than the apparent time.  (This is just one
example.)  This is true for all full virtualization schemes, not just z.

Recent Linux kernels for "System z" try to address this weirdness.
(Dunno if those updates affect Linux on other platforms.)  The same
problem appears for VMware hosting.  I would think it's true for Xen
also, but I haven't dived deep into Xen yet.  (Only have one Xen
running at home so far.)

Side note for perspective:

The IBM mainframe is better about virtualizing itself
than any other contemporary hardware.  Not completely sure why
but two features are strong suspects:  SIE and the I/O subsystem.
SIE is "Start Interpretive Execution", a misleading name for a
virtualization boost added some twenty years ago.  The I/O facilities
of all IBM class mainframes (even some workalikes such as most Crays
and some of the more successful Burroughs and Sperrys) is really sweet.
For I/O, the CPU points to a "channel program" and tells the I/O
subsystem to run it, as if to say, "handle that and don't bother me
until you're done."  I BELIEVE that these features let the hypervisor
(z/VM) do a better job of slicing up real resources.

Matters like the timing thing can still affect z/VM virtual machines.

Summary:

The real benefit of Linux on a mainframe is in large numbers.
You can have a handful of monstrously large Linux systems without z/VM
or you can have a boatload of normal sized Linux guest systems on z/VM.
The latter is what Nationwide and most mainframe Linux shops are doing.
(Means you have to jump into the virtualization game.)

It's great to have Linux on two or even three or more HW platforms.
You can then move workloads to where they run most effectively.
A re-compile is usually required, but "Linux is Linux" so that
everything else about the system is the same.  Only the instruction set
is different.

-- R;   <><