[Gllug] Crashes (owing to the heatwave?)

[Christian Smith]

On 8 Aug 2003, Rich Walker wrote:

>David Damerell <damerell at chiark.greenend.org.uk> writes:
>
>> On Friday, 8 Aug 2003, Chris Bell wrote:
>> >   Rows of expensive cabinets give no indication of what is inside a modern
>> >mainframe. I assume that a real computer does not have rows of risc chips in
>> >lots of packaging that make them appear slightly better than rows of 486's.
>> >   Do they use specialist chips, or groups of slightly better known chips
>> >working in parallel to get longer words?
>> 
>> The S/390 uses the IBM/Motorola PowerPC CPUs, which are of course also
>> found in Macintoshes (albeit not the exact same parts) and
>> conventional memory.
>
>Uses the POWER4 which comes as a bunch of multi-chip modules with
>insane cooling - you get 1 or 2 CPU's per module, with a big 128MB L3
>cache for each group of 4 or 8 CPUs
>
>The reason for "1 or 2" CPU's is that the high-performance computing
>crowd want memory bandwidth, and 2 cpu's in one module means lower
>bandwidth.
>

IFAIK.

S/390 (or zSeries as they are now) are neither PowerPC or POWER4.

They have their own architecture that goes back to the origional S/360
from the sixties, though extended to 64bits now. That is a different
architecture can be confirmed from the linux kernel assembly.

Hot plug in mainframes generally works on redundancy of CPUs. CPUs work is
checkpointed, and a CPU failure results in the last checkpoint being moved
to another, spare CPU, which picks up the work from that checkpoint.  
It's like database logging for CPUs. Generally, you can pull a CPU, then
plug it back in and it'll continue where it left. This is generally a
hardware feature, but there must be some software component to generate
alarms for failed CPUs etc.

Info:
http://www.research.ibm.com/journal/rd/435/spainhower.html

Linux hotplug CPU support is, I believe, software based in that the CPU 
must be taken offline before it is removed, else whatever is running on 
that CPU will be lost (and you'll be lucky not to panic the OS.) The 
benefit of this approach is that it is cheaper than fully redundant CPUs, 
and allows maintainance of the more common CPU issues such as CPU cooling 
problems or upgrades, for which there is warning of problems etc.

Linux 2.6 will also support hotplug RAM, I believe, in much the same way 
as hotplug CPUs work (ie. take it offline first!)

IBM partitioning works like VMWare, in that it provides virtual machine
resources and traps hardware access and emulates them. It is, however,
more efficient than VMWare, as the capability has been designed into the
hardware from the start, whereas x86 doesn't support virtual machines
without ugly software hacks. It isn't like partitioning used in the likes
of HP and IBM UNIX machines, as physical resourses are not fixed to
partitions (well, they can be, don't have to be.)

Cheers,
Christian

-- 
"You have not converted a man because you have silenced him." -John Morley

----------------------------------------------------------------------
    /"\ 
    \ /    ASCII RIBBON CAMPAIGN - AGAINST HTML MAIL 
     X                           - AGAINST MS ATTACHMENTS
    / \ 


-- 
Gllug mailing list  -  Gllug at linux.co.uk
http://list.ftech.net/mailman/listinfo/gllug