User: Password:
|
|
Subscribe / Log in / New account

Better per-CPU variables

Better per-CPU variables

Posted Nov 17, 2007 11:59 UTC (Sat) by ms (subscriber, #41272)
Parent article: Better per-CPU variables

"One of the great advantages of multiprocessor computers is the fact that main memory is
available to all processors on the system."

Hah!
1) It's simply stupid to write software that assumes this as it won't scale to distributed
systems.
2) If we didn't have shared mutable memory, vast classes of bugs, typically concurrency
related, wouldn't exist.
3) The direction in which CPU architectures are moving makes it less and less likely that a
single unified address space and general shared mutable memory are a) likely and b) desirable.

Hence the endless rise of languages like Erlang and Haskell which aren't built, unlike, say C,
on an unachievable model of a computer.


(Log in to post comments)

Better per-CPU variables

Posted Nov 19, 2007 1:56 UTC (Mon) by giraffedata (subscriber, #1954) [Link]

One of the great advantages of multiprocessor computers is the fact that main memory is available to all processors on the system.

That's like saying one of the great advantages of a lamp is that it illuminates things. It's not just a great advantage, it's the definition. What would make a system in which the various processors have separate memory a multiprocessor system?

1) It's simply stupid to write software that assumes this as it won't scale to distributed systems.

It sounds like you're saying it's stupid to write software for multiprocessor systems because then it would be software for multiprocessor systems.

I'm actually a foe of multiprocessor systems. My experience makes me believe processors separated by TCP/IP, FCP, etc. links are more profitable for most things. But I can definitely see the advantages of SMP.

Better per-CPU variables

Posted Nov 22, 2007 7:52 UTC (Thu) by eduperez (guest, #11232) [Link]

One of the great advantages of multiprocessor computers is the fact that main memory is available to all processors on the system.

> That's like saying one of the great advantages of a lamp is that it illuminates things. It's not just a great advantage, it's the definition. What would make a system in which the various processors have separate memory a multiprocessor system?


I think ms was talking about NUMA architectures.

Better per-CPU variables

Posted Nov 23, 2007 6:28 UTC (Fri) by giraffedata (subscriber, #1954) [Link]

I think ms was talking about NUMA architectures.

Everything ms says, and everything the article says, is applicable to to all multiprocessor systems, NUMA and non-NUMA alike. They all have shared memory and they all have difficulty because of that.

NUMA shares memory in a way that strikes a different balance between the difficulties and the advantages of shared memory than other multiprocessor systems, but it remains a defining characteristic, not just an incidental feature, of multiprocessor systems that they have shared memory.


Copyright © 2017, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds