[Info-vax] VMS on Raspberry Pi 5

[Dan Cross]

In article <1f49caa7-6d19-4946-a433-3e85492b3e02n at googlegroups.com>,
Jake Hamby (Solid State Jake) <jake.hamby at gmail.com> wrote:
>On Thursday, November 16, 2023 at 5:36:21 AM UTC-8, Scott Dorsey wrote:
>> Ahem A Rivet's Shot <ste... at eircom.net> wrote: 
>> > 
>> > Yes they are popular with banks and the like because they *also* 
>> >run their old OS-360 stuff without recompiling it, but to anyone who 
>> >doesn't need that they are very expensive for little gain. Guess what many 
>> >of their customers run in the RHEL environments - yep kubernetes and docker.
>> The IBM systems are I/O machines. The CPU is just sitting there telling the 
>> I/O controllers what to do and most of the real work is being done by other 
>> hardware outside the CPU. So you can have incredibly high workloads and 
>> huge transation rates with relatively slow CPUs. 
>
>This was definitely the case 20 years ago. Today, IBM Z CPUs
>are quite interesting in their own right. Massive caches. The
>previous z15 CPU had four levels of cache, with a whopping 960
>MB of L4 cache (eDRAM) per CPU drawer. The current z16 has a
>different architecture with 32 MB of L2 cache per core and an
>innovative cache design that allows any CPU to borrow cache RAM
>from other CPUs on the same die, different dies, or even
>different CPU drawers, dynamically creating arbitrarily large
>virtual L3 and L4 caches.
>
>They're never going to beat x86, ARM, or POWER on raw per-core
>compute speed, but for the giant databases their customers run,
>with terabytes of RAM for caching and massive I/O throughput,
>mainframe CPUs are quite fast. They have SIMD for decimal
>floating-point (as well as binary FP and integer math) to
>accelerate COBOL, an embedded sorting engine for in-memory
>"DFSORT" acceleration (sort/merge/copy), an embedded zlib
>engine for fast compression, and now an embedded neural engine
>to enable running AI fraud detection models on every
>transaction.

There was an interesting talk at MIT a few months ago about how
specialized accelerator will be more or less required to
continue scaling.  IBM has obviously embraced this on the
mainframe, but they've been using specialized satellite
processors to offload lots of processing from the main CPUs for
decades, as has been mentioned.

>IBM's sales pitch for running Linux on Z is that it's more
>power-efficient to have one or two mainframes than a much
>larger number of x86 blades doing the same thing. I'm sure that
>depends very much on workload. I should also mention z/VM,
>which is IBM's OS dedicated to running transactions (as opposed
>to z/OS, which is more general-purpose). z/VM is what all the
>airlines and Visa credit card transactions run on.

I find this surprising.  z/VM is a hypervisor, mostly for
running VMs: users either log into a VM running under z/VM
or boot CMS (a small single-user operating system) that they
interact with.  I think of it more as a vehicle for running
other systems that do useful things, like z/OS, CICS, and
all of that stuff.

>Finally, there's Parallel Sysplex, which is basically
>VMSclusters with dedicated hardware acceleration. Between that
>and all the redundancy and other RAS features, IBM can tout
>independent surveys showing 8 or even 9 9's of uptime! Their
>POWER10 servers don't look too shabby either.
>
>https://techchannel.com/Trends/07/2023/ibm-z16-power10-high-reliability-servers
>
>Of course you're paying a premium for all that custom
>engineering, but I think it's impressive, nonetheless.

Honestly, I think top-end x86 gear gives mainframes a run for
their money at this point.

	- Dan C.