It's most likely to do with how 
ioRelinquishProcessorForMicroseconds(int microSeconds) handles 
sleeping.

When there is no Smalltalk processes able to run, the UI is waiting 
for a UI event, all Delays are waiting to expire and all the sockets 
are waiting on input the Smalltalk Scheduler schedules the lowest 
priority process which asks this primitive to wait a few microseconds. 
This is somewhat bogus since we ask to sleep a few microseconds.

So a few years back I recommended altering things to wait until the 
first pending Delay finishes or if a socket/ui event requires 
servicing, that would maximize sleep and I'd hope most mac  users 
would say their Squeak VM uses under 10% when idle.  {Comments are 
welcome on this}

I'll explain how the os-x carbon one works, I can't comment on the 
windows code, but as you've seen it's not sleeping as much as expected.

The os-x carbon code does: 

    /* Wakeup interrupt processing once we return because there will be  
a pending interrupt we'll need to service */
    setInterruptCheckCounter(0);
    /* get the time now, and compare to the time we need to wake up if  
we have a pending Delay running.
   If there is not pending delay then we'll sleep 16 milliseconds. With  
MVC this was true, with Morphic this is never true.
   If the wakeup time in is the past, just return don't sleep */
    now = (ioMSecs() & 536870911);
     if (getNextWakeupTick() <= now)
         if (getNextWakeupTick() == 0)
             realTimeToWait = 16;
         else {
             return 0;
     }
     else
         realTimeToWait = getNextWakeupTick() - now;
    /* Call this unix async poll routine to nano-sleep until we exceed  
the time to wait,
 or if a user generated event or socket event occurs. This ensures we  
maximize sleeping, however
 morphic drives a UI poll event every 1/60 of a second, still we'll get  
only 6ish% CPU when idle... */
     aioPoll(realTimeToWait*1000);

If you have access to the windows C code and a desire to build your 
own VM then it may be possible to reduce the CPU usage if you have a 
good understanding of the side-effects to making that change to 
ioRelinquishProcessorForMicroseconds.

On Sep 17, 2004, at 4:01 AM, Dmitrij Abolichin wrote:
> Hi,
> We have small app that runs on server and acts as kind of "task
> runner".
>  Squeak is running as Windows service, and consumes 50% of CPU
> resources. At that time Squeak CPUWatcher indicates, that "idle
> process" consumes about 100%.
>  Can anybody explain why Squeak.exe takes so many CPU resources?
> VM version 3.6.2, Squeak ver. 3.6-5424