Bonita Montero <Bonita.Montero@gmail.com> writes:
>Am 17.05.2023 um 17:11 schrieb Scott Lurndal:
>
>> Bonita Montero <Bonita.Montero@gmail.com> writes:
>
>>> I gave hin an example of how eficient I/O is on a 1.500$ PC.
>
>> Apples are red, oranges are, well, orange.
>> How well did your PC handle I/O in 1975?
>
>There are still mainframes and they're
>for sure not more efficient with I/O.

And you know this, how, exactly?

How many disks can you attach to your $1500.00 PC?
(And why on earth would you spend that much for a PC?)

How many DMA transfers can your mainboard chipset
support simultaneously? What's the bisectional bandwidth
of your processor/memory interconnect fabric?

Now compare that to the latest Z-series processor family.

Am 17.05.2023 um 17:37 schrieb Scott Lurndal:

> And you know this, how, exactly?
I/O on mainframes is also PCI-whatever today.

On 17/05/2023 16:44, Bonita Montero wrote:
> Am 17.05.2023 um 17:37 schrieb Scott Lurndal:
>
>> And you know this, how, exactly?
> I/O on mainframes is also PCI-whatever today.
>
Back when I worked on mainframes, which might be before you were born,
some of the big systems had multiple memory units each of which could be
connected to disc and/or tape drives.

They were of course all connected to the CPU cluster.

Going to a PCI based system would tend to indicate that there is a
single bus taking all the I/O and CPU traffic, which would be a bottleneck.

If the mainframes of 40 years ago had got around that bottleneck why
would it have been introduced in the time since?

I suspect others here have more recent knowledge than I do.

Andy

On 5/23/23 6:33 AM, Vir Campestris wrote:
> On 17/05/2023 16:44, Bonita Montero wrote:
>> Am 17.05.2023 um 17:37 schrieb Scott Lurndal:
>>
>>> And you know this, how, exactly?
>> I/O on mainframes is also PCI-whatever today.
>>
> Back when I worked on mainframes, which might be before you were born,
> some of the big systems had multiple memory units each of which could be
> connected to disc and/or tape drives.
>
> They were of course all connected to the CPU cluster.
>
> Going to a PCI based system would tend to indicate that there is a
> single bus taking all the I/O and CPU traffic, which would be a bottleneck.
>
> If the mainframes of 40 years ago had got around that bottleneck why
> would it have been introduced in the time since?
>
> I suspect others here have more recent knowledge than I do.
>
> Andy

I think part of the answer was that mainframes were much more of a
"Distributed" system then more modern machines. The "I/O Channels" were
very much a form of processor assigned to do certain pieces of the work.

The logic was slower, so the wires between the modules wasn't as
critical, so you could have fatter connections.

Vir Campestris <vir.campestris@invalid.invalid> writes:
>On 17/05/2023 16:44, Bonita Montero wrote:
>> Am 17.05.2023 um 17:37 schrieb Scott Lurndal:
>>
>>> And you know this, how, exactly?
>> I/O on mainframes is also PCI-whatever today.
>>
>Back when I worked on mainframes, which might be before you were born,
>some of the big systems had multiple memory units each of which could be
>connected to disc and/or tape drives.
>
>They were of course all connected to the CPU cluster.
>
>Going to a PCI based system would tend to indicate that there is a
>single bus taking all the I/O and CPU traffic, which would be a bottleneck.

In the mainframes, there was still a single bus from the memory
subsystem to the I/O subystem. And in those days, 8MByte/second
was a high-speed transfer rate from memory to the I/O controller,
and 1Mbyte/sec was the typical transfer rate between a disk
drive and the disk controller, allowing eight controllers to
simultaneously transfer data to/from memory (interleaved with
processor accesses to the same memory).

The PCI bus designed in the late 80s/early 90s was indeed a bus,
(originally at 33mhz, 66mhz for PCI-X).

Modern PCI express is _not_ a bus. It's a point-to-point serial link
between the host memory complex and an I/O controller (storage
host adapter, network interface card, etc).

PCI Express Generation 6 will transfer 128 Gigabytes/second
in a 16-lane configuration, and most high-end processors support
multiple PCIe 6.0 controllers.

>I suspect others here have more recent knowledge than I do.

See above.

On Tue, 23 May 2023 14:51:25 GMT
scott@slp53.sl.home (Scott Lurndal) wrote:
>Modern PCI express is _not_ a bus. It's a point-to-point serial link

Though these days "serial" means little more than a single data line. Modern
serial comms such as USB and SATA are packet based and have more in common with
ethernet than RS232.

Am 23.05.2023 um 12:33 schrieb Vir Campestris:

> If the mainframes of 40 years ago had got around that bottleneck
> why would it have been introduced in the time since?

Whatever PCIe I/O-devices mainframes have today - I get about ~1.1E6
/ ~4GB/s random 4kB-I/Os per second on my 90$ Samsung 980 Pro 1TB PCIe
4.0 SSD with 3 - 4% CPU load (AMD 7950X, I/O completion ports - Perfor-
mance Monitor shows ~1.1e6 interrupts per second) under Windows 11 and
I guess that today's mainframes do that not much more efficient.

On 23/05/2023 15:51, Scott Lurndal wrote:
> In the mainframes, there was still a single bus from the memory
> subsystem to the I/O subystem. And in those days, 8MByte/second
> was a high-speed transfer rate from memory to the I/O controller,
> and 1Mbyte/sec was the typical transfer rate between a disk
> drive and the disk controller, allowing eight controllers to
> simultaneously transfer data to/from memory (interleaved with
> processor accesses to the same memory).

My experience was ICL 2900s, and AIUI they could have multiple memory
subsystems. Moving data between them was something we did our ****est to
avoid.

Andy

On 5/16/2023 9:56 PM, Chris M. Thomasson wrote:
> On 5/16/2023 9:22 PM, Bonita Montero wrote:
>> Am 17.05.2023 um 05:24 schrieb Chris M. Thomasson:
>>
>>> Be careful with those IO completion ports for they use non-paged memory!
>>
>> The kernel uses non-paged memory for a lot of data structures.
>>
>
> You need to be weary about how many in progress IOCP actions there are
> in the system at the same time. They will eat into your non-paged pool.

This of stress testing a system until it blows up with IOCP...

https://youtu.be/HEYrRNMr2kg?list=RDMM