On 2/20/07, Tom Duerbusch <[EMAIL PROTECTED]> wrote:

In what cases is this a problem?  I want to know, in what circumstances
I'm going down a path, that is different than I think.

I fear I don't know the details anymore of that particular event. All
I remember is that fixes were required to make Linux use mini disks
allocated on a later or different model of the ESS. But if you look at
a snippet of source code like here for example
http://lxr.linux.no/source/drivers/s390/block/dasd_eckd.c?v=2.6.18;a=s390#L67
you can see the information is there and there obviously was a reason
to differentiate.

And I certainly know about one case where disk I/O performance was
worse on new DASD because of such differentiation. Would you believe
that still in this century our friends in Boeblingen implemented RPS
in the disk driver?

Linux is not unique in this. I still remember that the SE of our
non-IBM DASD implemented (unplanned) a performance fix for the
microcode. All it did (as we learned after brown stuff some air
movement devices) was to set a bit in some status field returned to
the OS. And z/OS used that to identify features of 3990 control units.
While the concept did not apply to this RAID device, it fooled z/OS
enough to exploit some of the features. Unfortunately, VM passed that
status field also to VSE guests that then failed to recognize the
control unit and gave up.

For example, right now, our zLinux disaster recovery option is to find
a replacement processor (perhaps the next processor off of IBM's line),

Would you also be willing to take the load of a z990 to someone's z9
for disaster recovery? Back then when I was involved in such things I
do remember we applied preventive maintenance on VM for future models
because of such arrangements. But VM does not isolate the guest from
those details (as folks using ISV products with CPUID based license
code know). We just have been spoiled by CMS behaving so well that we
take this all for granted.

I often disagree with Alan (also) about this. Probably because I
sometimes do get out ;-)
I believe VM should reveal far less details about the actual resources
so we don't challenge developers to take advantage of the details that
shine through our virtualization. I believe there's often no value for
a class G user to know what real device or real volser his mini disk
is - it's a disk and it's so big, and that's what you need to know.
Let alone a Q MDISK that reveals the actual owner for indirect links.
Yes, we had a smart developer who found which disk $FORTRAN 200 really
was and improved his performance by linking that directly. So after I
upgraded he still used the old version and reported problems that I
was sure to have fixed...

If we did not have VSWITCH, most certainly the OSA devices would have
caused a nightmare too. The problems we have seen there are mostly
because they chose to simulate existing hardware to avoid writing one
extra clean and elegant driver in Linux.

It all comes down to abstraction and architecture. VM needs to provide
an abstraction of the true resources in such a way that the virtual
machine can make clear what function is needed, and CP can then
deliver that in the "best" way (weighing technical and economical
aspects). The challenge is to have an abstraction that is powerful and
flexible enough, and yet efficient to implement and compatible till we
retire. Normally that means a high-level interface. To have both VM
and the guest imitate a low-level interface of a device that was
extinct already 25 years ago, that's a burden to both...

To define such abstractions is hard work, but I strongly believe it is
more effective than trying to imitate the latest hardware. If it were
my choice, Linux on z/VM would just be using (an improved version of)
the diagnose I/O interface rather than doing channel programs.

Yes, I will shut up now. Rob

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390

Reply via email to