On the Linux systems here all cores are run at close to 100% CPU, then the CPU frequency reduced or increased as necessary to run the CPU at an almost constant pre-set temperature - this works for Summer and Winter

To measure heat of the CPU, a free utility this can help:
https://openhardwaremonitor.org/

For me, my i5 laptop can only run 75% 50% of the time without getting temps above 70°C under normal workload. I imagine it differs based on your computer.



Do know what your RAC or credits given on those machines? On my i5-3470 CPU @ 3.20GHz (4 processors), my RAC is about 1600 (~15 GFLOPS?)

Your T7300 is ~5 GFLOP/s according to this: https://boinc.bakerlab.org/rosetta/cpu_list.php

My kill-a-watt says 60W. So, for the same energy usage, I'm getting about triple the crunching.

I think upgrading to newer hardware is the best option so your energy to RAC ratio is much better. To put it another way, you could turn off a newer computer all summer and still crunch more than two of Core2Duo computers the rest of the year.

You can buy an Dell OptiPlex 3010 on eBay for under $50 USD.

I've found controlling the CPU frequency to be the best way to go because it keeps everything steady. It limits power fluctuations and the fan speed doesn't change often like it does if you are using a temp throttling or CPU time approach. I don't have any data on this, but it seems like this would result in less wear and tear on the relatively fragile laptop components. I'm not sure which would give you better results between using only 1 core at 100% (or 99% to disable any CPU boosting) vs. using both cores at 50% frequency. I think either of those options is much better than using both cores at 50% CPU time.

I agree with the above comments, avoid thermal throttling, go for CPU cycle control. Its a basic rule of electrical engineering that stopping and starting any DC motor (ie computer cooling fan) causes high inrush currents (up to 20 times normal running current) and over time this shortens the lifetime of the motor, and could marginally increase your power bill, but if the throttling is happening in cycles of less than a second the fan would not have time to stop, and the effect would probably be negligible. What you want is smooth, uniform current and heat output, for less overall heating-cooling cycles, because such cycles eventually create thermal stress on all the components. However, setting a thermal throttle with a slightly higher temp setting than say, 100% CPU on a hot summer day is a good idea just in case the cooling fan fails, or as sometimes happens, starts to become inefficient due to being clogged with dust.

You can throttle very easily on any modern Linux using systemd v231 or higher (for example, I use Debian 9) to get where you want, I do this on all my nodes for the same reasons in this thread. It depends on the specific CPU, but "roughly 60%" seems to be the point where most seem to like it. At this level the fan just barely turns on (~3000-ish RPMs) - I measure a 22W draw (laptop battery, using lm_sensors) at 60%-ish for an i5-3320m (Dell E6330).

First, calculate what 60% means to your number of cores/threads being used in aggregate - systemd considers each "visible CPU" as 100%, so a dual-core hyperthread-on CPU comes out to 400%; "60% of each visible CPU" is therefore 240% (400*0.6) to systemd's needs.

Next, set the throttle in place on the boinc client:

EDITOR=vi systemctl edit boinc-client

# Add the below to the new file being edited and save
[Service]
CPUQuota=240%

Re-read the new design into systemd's running brain and restart the boinc client:

systemctl daemon-reload
systemctl restart boinc-client

Now set the boinc preferences to run in always-on mode and to ignore "activity" (because we're throttling it outside it's world view instead):

vi /var/lib/boinc-client/global_prefs_override.xml

<global_preferences>
  <suspend_cpu_usage>100.000000</suspend_cpu_usage>
  <cpu_usage_limit>100.000000</cpu_usage_limit>
  <leave_apps_in_memory>1</leave_apps_in_memory>
  <run_if_user_active>1</run_if_user_active>
  <run_on_batteries>1</run_on_batteries>
  <run_gpu_if_user_active>1</run_gpu_if_user_active>
</global_preferences>

boinccmd --read_global_prefs_override

If you had running work units, I would now go into the client and run the CPU benchmarks again (I like using "boinctui" - nice, simple client) under their new limits just to true up all your estimated timing stats. It runs periodically on it's own though so if you forget to do this, no big deal.

You can then play with this to up or down the percentage to fine tune it to your exact CPU and cooling situation, I like the package "lm_sensors" for this monitoring as it easily shows all the temperatures and fan speeds - and if it's a laptop, it can measure the power draw while on battery (to calculate your cost per month).