Optimal Win2K SMP config?

[tain]

Tomcat76 nearly has me convinced to buy a new rig like his but in the process of debating several issues I found myself wishing there were a faq/guide about optimal SMP configurations for Windows 2000 Pro. Perhaps this thread will serve that purpose if such a guide doesn't exist.

As I understand it, there are several factors to consider as they specifically relate to Windows 2000:

• Win2K Pro supports two processors.
  
• Performance of two single core processors vs. one dual core processor.
  
• Win2K doesn't properly recognize dual core processors, but can fully utilize both cores depending on the CPU memory scheme:
  • How does shared vs. separate on-die memory impact Win2K support of dual cores?
    
  • AMD vs. Intel on-die memory implementations (shared vs. separate) in dual core chips.
    

  [*]Given limited hardware availability, is a four-core Windows 2000 Server (with four processors or two dual core processors) prudent at this time?

Based on most of our research so far, it appears that AMD's FX, Opteron and X2 (preferably socket 939) are the only dual core processors that Win2K will recognize and fully utilize, with the X2 939 seeming to be the "sweet spot" of price vs. performance. Intel doesn't appear to be much of a player here since their dual core chips use shared memory which Win2K can use, but not for SMP.

I am also interested in a "fast, but budget SMP" config versus an "all out, fastest Win2K SMP config ever because future hardware won't be properly recognized" config, if that point has been reached.

Note: This thread has nothing to do with overclocking.

Thanks to Tomcat76 for all his help with this!

[tain]

A very knowledgable friend of mine provided some comments on this topic and I thought somebody out there might appreciate them.

The problem with SMP on 2K is that it doesn't understand that a dual-core shares the same L2 between cores.

HyperThreading works like absolute s***e, because it treats the HT as a real second proc - the HT shares both L1 and L2, and causes horrible cache-thrashing. So I would never run HT under 2k.

As for dual core, you end up swapping L2 out to RAM on a busy box frequently enough that you notice the difference between a real dual proc box and a dual core box on 2K. On 2K3, they're pretty close because the kernel understands that both cores share an L2, and tries to run threads in a manner that effectively utilizes cache better.

The reason 2K runs fast is because it's basically NT4 underneath (at the kernel) and that runs fast, compared to 2K3. But there's a reason 2K3 runs better - all that extra "bloat" people complain about makes the box run *smarter*.

The reason 2K3 runs a little slower as a desktop OS is because it's quantum is much longer than XPs, meaning threads can run longer on 2K3 before they're swapped back out to the wait list even if you choose "programs" over "cache". You can tweak it in the registry, but then you can just use XP unless you need the incoming net connections

[tain]

Also, tain, I had a look at what your friend wrote... in some respects, I think he might be confused.

Windows doesn't need to know that an L2 is shared between cores. If you think about the function of an L2 cache, why would it? No version of Windows gets "told" by the processor the L2 configuration, and no version of Windows has any use for "knowing" the configuration. Windows cannot "run threads" in a manner fundamentally differently based on the configuration of L2 -- L2 is the waiting room before instructions are executed by the processor. L2's just cache RAM, and it's the size of it that the OS needs to know and 2003 autodetects. Windows 2000 and XP can be set manually (Google search).

Also, Hyperthreading has nothing to do with the Core 2 Duo and Quad chips (maybe he knows this). The Core 2 chips have dropped Hyperthreading support.

Windows 2003 runs better because it makes the box run "smarter?" So, running Fax Services is running smarter? Running IE Dlls at all times is running smarter? The only thing this holds true for is Active Directory. Have your friend uncompress and load SYSSETUP.INF and look under [infs.Always].

If all of those INFs make 2003 run "smarter," then I'm a n00b and a monkey's uncle. 2003 is bogged down because it's bogged down, and if he doesn't believe it, he needs to expand and have a look inside some of those INF files listed under [infs.Always]. What he says about execution quanta happens to be correct.

Bottom line -- 2000 Server will see 4 processors on a Core 2 Quad, and it's my aim to take advantage of that once I get a finalized fileset.

I have friends that are so much smarter than me. I'll just keep posting their thoughts here