[Nomen-dev] Cost of synchronization

[Brian H. <bh...@sp...>]

Hmm.  Just ran a curiosity test.  Benchmark machine was a 1.4GHz P4, ~1% 
CPU utilization, Linux 2.4.  I tested several things (the power of 
ifdefs):

- An empty loop, to see how much overhead the loop itself took
- An xorl $1, memory instruction
- the above insturction with a lock prefix
- an xchgl reg, memory instruction (which implicitly asserts LOCK#)

The empty loop did one loop every about 1.5 clock cycles (this was gcc
-O6, but I looked at the assembly language to make sure the loop was still
there).  The non-locked xorl took 15.3 clocks per loop, the locked xorl
took 125.6 clocks per loop, the xchgl took 127.4 clocks per loop.

You can throw a lock prefix on a usermode instruction, but it adds about
110 clocks onto the cost of the instruction- painfull!  Lock elimination
is going to be a very worthwhile endeavor.

Brian