Page 2 of 2
Re: Usecase idea.
Posted: Fri Nov 20, 2020 1:24 pm
by nexos
eekee wrote:nexos wrote:I'm thinking of a server with 100s of CPUs and a couple terabytes of RAM. It could even be a distributed system. It would be cheaper probably (less computers, just a few very pricy computers).
Okay. I know nothing about how such a system handles ram contention, so I'll leave it at that. It's funny you should mention distributed systems. Before the term was defined as narrowly as it is today, Plan 9 was called a distributed system. I realise now I've forgotten all about the details of distributed systems.
The answer to RAM contention is NUMA. NUMA allows each CPU to have its own private memory that it has fast access to. It can access other memory, but that is slower. An OS should be NUMA optimized so it can take advantage of the extra speed of NUMA systems.
Re: Usecase idea.
Posted: Fri Nov 20, 2020 4:54 pm
by Korona
NUMA systems do not bring extra speed, they just avoid extra slowdown.
Re: Usecase idea.
Posted: Fri Nov 20, 2020 5:05 pm
by pvc
Didn't we stray from original question a little bit too much?
Re: Usecase idea.
Posted: Fri Nov 20, 2020 5:32 pm
by Octocontrabass
Korona wrote:NUMA systems do not bring extra speed, they just avoid extra slowdown.
...And that's only if you optimize them correctly. Linux, for example, performs notoriously poorly when a single program exhausts one CPU's available resources. Linux database configuration guides typically recommend disabling NUMA optimization.
pvc wrote:Didn't we stray from original question a little bit too much?
That depends on how many CPUs you need for all of your users.
Re: Usecase idea.
Posted: Fri Nov 20, 2020 6:26 pm
by pvc
I was thinking about something like Threadripper class machines. 64 cores now, but maybe something like 256 or 1024 in about 5 years.
Re: Usecase idea.
Posted: Sat Nov 21, 2020 8:11 am
by eekee
nexos wrote:The answer to RAM contention is NUMA. NUMA allows each CPU to have its own private memory that it has fast access to. It can access other memory, but that is slower. An OS should be NUMA optimized so it can take advantage of the extra speed of NUMA systems.
Thanks, that makes a lot of sense. (I remembered the term and should have looked it up.) It sounds like a NUMA system is somewhere between multiblade and original SMP; that's really cool.
@pvc: I was trying to question the performance of the original idea. My questions have now been answered to my satisfaction, so I look forward to seeing it develop.