Re: Threads across programming languages

On 5/4/2024 10:25 AM, Ross Finlayson wrote:

On 05/04/2024 09:03 AM, Ross Finlayson wrote:

On 05/03/2024 09:34 PM, Chris M. Thomasson wrote:

On 5/3/2024 9:30 PM, Chris M. Thomasson wrote:

On 4/29/2024 9:44 AM, Ross Finlayson wrote:
[...]
>
Have you ever read the following paper? Pretty nice!
>
https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-2001-39.pdf
>
>
>
If not, read all...

>
There is a way to do a little sort, gain a cohort and execute it. For
instance all reads, writes, connections, accepts are all organized in a
cohort.
>
Instead of processing a read, write, read, read, write. Say:
>
read, read, read, write, write
>
To get better locality. I remember some functions on winnt. AcceptEx,
ConnectEx, TransmitFile, TransmitPackets, ect... ;^)
>
The main thing to use to wait on io completions:
>
https://learn.microsoft.com/en-us/windows/win32/fileio/getqueuedcompletionstatusex-func
>
>
>
https://learn.microsoft.com/en-us/windows/win32/api/mswsock/nf-mswsock-acceptex
>
>
>
https://learn.microsoft.com/en-us/windows/win32/api/mswsock/nc-mswsock-lpfn_connectex
>
>
>
https://learn.microsoft.com/en-us/windows/win32/api/mswsock/nf-mswsock-transmitfile
>
>
>
https://learn.microsoft.com/en-us/windows/win32/api/mswsock/nc-mswsock-lpfn_transmitpackets
>
>
>
I used to use them all the time back in the day.

>
How about Hillis' "Connection Machine" and xectors.
>
"Affinity" is the most usual sort of notion of
locality and reference and that things that are
local are simpler to address and keep coherent,
having smaller address offsets.
>
SIMD or single-instruction-multiple-data and
"clubbing cohorts" of routines that have a
common program counter in their next instruction,
seems a pretty simple idea, what with what enables
this is usually "content associative memory",
that routes things into buckets rather naturally.
>
(In silico.)
>
It's like "we'll just use content-associative-memory
and get a huge speed-up" and it's like "oh, what
if it's otherwise occupied". It's like "oh we'll
just add burst-buffers everywhere between units of
the traces on our dies to modulate rate control"
and it's like "there's no room on the layout".
>
Then it's like "the chips made this and we wrote
a scheduler and now it's according to us" and
it's like "yeah, in a sense, it's system programming".
>
>
Which of course is quite "old hat" and all has "prior art".
>
>

  It reminds me of this time about fifteen years ago,
I was at my desk and the phone rang. I answered
the phone to "This is monitoring. Are you running
a script?" And I say "Yeah, I'm running a script on
monitoring." And they say "It's degrading prod."
I say "Do you want me to stop it?" And they say "Yes."
So, I hit Ctrl-C in that terminal as I'd designed my script
to populate all the monitors and alerts and alarms for
all the metrics of all the APIs and granular to the client
and with rollups and suppressions to distinguish outages
from use-cases, across all services across all regions,
to not much care. Then I added a brief pause between
calls and re-launched it.
 Sometimes using all the resources isn't very fair.
 

Think of a client that would try to make the server use all of its resources and actually get into a major danger zone, even flag a malloc returning zero. Then there is non-paged memory to consider. On windows, each in flight OVERLAPPED operation uses non-paged memory... If that goes down, well, its not good at all... ;^o