It's not measured in anything comparable to real computers. Emulation counts cycles (i.e., instructions per msec) much differently than real CPUs do. Your target should be somewhere between 10000 and 25000 cycles, but that varies with different games depending on what instructions they use. These numbers are simply a value from my own experience, there's nothing calculated.

1000 dosbox cycles seems roughly equivalent to 3-6Mhz on a 486DX, as a very general rule, but like Moe says it depends a lot on what you're running. It is possible for emulators to work on a clock rate (Mhz) basis rather than number of instructions/second but it makes emulation slower and more complicated, so DosBox doesn't do this (afaik).

Is there any way to practically benchmark DosBox performance?

I ask this because increasing cycles always fools the application that's run in DosBox that it's running faster, even if it's not.

I'd be interested to know if there's objective, measured, way of doing this instead of by 'hmm..this seems a bit slower, perhaps I'll drop the cycles by 100'-gut feeling.

You can try good old Norton SI. Older versions measured speed relative to XT. 1000 cycles gets me 8.2 XTs on my system. IIRC, my 486DX/33 was ~32 XTs, I need around 4000 cycles to get that. My system is AMD64 3500+ somewhat slowed down by non-optimal RAM config.

Doesn't that give *in emulation* speed where more cycles always means faster emulation?

What I'd like to know is the real emulation speed: I have no idea how many cycles my comp can handle before adding more cycles becomes harmful.

Oh, if that's all you want: Turn them up until your sound gets choppy. Then turn cycles down a little bit until sound is OK again. That's the optimal setting. Different games have different settings!

By the way: Too high cycles will never harm your computer, it just means that DOSBox will run slower and with choppy sound.