sidegrade? upgrade? Cores or MHz?

eccerr0r · Posted: Sat Aug 26, 2023 3:21 pm Post subject: sidegrade? upgrade? Cores or MHz?

I recently got an ancient LGA2011 motherboard and now have the possibility to get a fairly large SMP CPU, up from 4 cores.

However it looks like I may get a frequency cut, though not a huge loss because turboboost will make up for some of it, but not all, and definitely less when the core count goes up.

So... what do you all think, should I take a frequency cut to get a core count bump (these are actual cores, not HT)?

The proposal:
From: 3.6GHz 4 Core

To: 3.5GHz 6 Core
3.3GHz 8 Core
3.0GHz 10 Core
2.7GHz 12 Core

Is it worth to drop almost 1GHz to get 3x the number of cores?

Assume I can get RAM to 2x the number of cores though it is not the case now, and of course the target application is: emerge (and its forked children) :-)

Note that the concern is that a lot of portage doesn't seem to parallelize and thus runs singlethreaded, which will bottleneck on a low clock rate core.

This experiment is totally curiosity of throughput increase over a 4C and distcc to other full complete machines which probably consumes more power ...

I suspect the threads will increase ram requirements further as these are all HT machines, so more ram will be considered, but for now just deal with the actual cores...

Opinions, anyone?
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?

Ralphred · l33t Joined: 31 Dec 2013 Posts: 653

I'm going to do simple* maths and ignore overhead and configure threads.

eccerr0r · Posted: Sat Aug 26, 2023 7:59 pm Post subject:

Indeed that is the simplest way to determine effective throughput for perfectly parallelizeable software.
Unfortunately the real world isn't like that and yes it's specifically the single thread programs like portage housekeeping that only use one core.
Plus there are sections of builds that just chew on one thread, leaving all other cores idle...

Yes the 8 is tempting, plus don't need to load up as much RAM... just that the 10 and 12 look impressive on 'top' ... anyone with massively parallelized machines and able to keep all threads occupied? Heck for 4(8 HT) it's easy to end up with one core/thread idle...
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?

Spanik · Veteran Joined: 12 Dec 2003 Posts: 1003 Location: Belgium

eccerr0r · Posted: Sun Aug 27, 2023 11:55 am Post subject:

Yep as I thought, perhaps 12 is not something I really want, and 8 or 10 would be more ideal (so I can keep turboboost going on when portage goes single threaded)... hmm.
True too that perhaps this machine should go to my VM host instead of being a workstation, though my VM host is currently fine as it is (it's also a 4 core machine that only runs 2 VMs.)

Technically speaking it's solely RAM that's hurting at the moment on my VM host, not core/cpu throughput hence not considering a 12-core for it. Still debating whether I should downgrade the old core2quad board to an 4C atom board - I lose a bit of speed but gain more RAM (ddr2 -> ddr3 as ddr3 dimms tend to hold more), but this is straying off topic :)
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?

Spanik · Veteran Joined: 12 Dec 2003 Posts: 1003 Location: Belgium

Of course YMMV. I have no idea what your typical workload is. Like you I noticed that portage is often stuck in single thread for a long time. Mostly during the pre-compile checks. And for some packages these take longer than compiling the lot. And my VM's are mostly running XP on 4 cores/4GB. At most 2 of those are running at the same time.
_________________
Expert in non-working solutions

logrusx · Advocate Joined: 22 Feb 2018 Posts: 2432

eccerr0r · Posted: Mon Aug 28, 2023 12:31 pm Post subject:

I've experienced pains with those virtuals and packages that don't have compile requirements on really slow machines, hence the desire for high clock rate machines for those unparallelizeable sections. Like on my 1.6GHz Pentium M an 1.6GHz Atom CPUs take 30 seconds to install a virtual/no compile package, this does add up to contribute to 5 day upgrade cycles (yes, the nearly 1-day builds of rust and gcc contribute to most of it, but that's just 2 of the 5 days over the hundreds of packages that are scheduled to merge).

---

Hmm... overclocked the cpu to 4.4GHz... now it makes the calculus harder...

---

Nope, math got a bit easier: 4.4GHz is not stable, 4.1GHz seems to run the benchmark I run that gets cpus real hot (it happens to be a java program).
_________________
Intel Core i7 2700K/Radeon R7 250/24GB DDR3/256GB SSD
What am I supposed watching?