In short we can definitely say that there is performance differences, but usually good programmers can optimize very well and get around it.

But to give some examples:

example a) you run 4 processes at a single cpu and they are mutually locking.

In windows you implement it for example with the well known WaitForSingleObject()/WaitForMultipleObject().

In Linux you lock as it has no WaitForSingleObject().

In IRIX you also lock using spin_lock().

Then you measure performance and linux is tens of times slower than both the other OSes.

It appears then that both IRIX and windows are first trying to lock the object a few times in a row without putting a process into the runqueue. So if such an object is released within say 600 ns then the processes can run on.

However under linux there is no such function for multiprocessing, so each process gets in the runqueue which means a 10ms latency each time, which is pretty much slower than < 600 ns :)

Trivially some good programming can speedup the software at all the OSes a lot.

Paul DeMone (pdemone@igs.net) on 6/22/03 wrote:
---------------------------
>Anonymous (nn@nn.invalid) on 6/22/03 wrote:
>---------------------------
>>Paul DeMone (pdemone@igs.net) on 6/20/03 wrote:
>>---------------------------
>>>The SPEC CPU suite programs are specifically selected and sometimes even
>>>modified to minimize the influence of differences in OS and language libraries
>>>across platforms.
>>
>>Then why is e.g. Intel's compiler for Linux showing noticeably lower performance
>>than Intel's compiler for Windows? Last evidenced in Ace's Hardware's posting about Opteron SPEC performance:
>>
>>http://www.aceshardware.com/read_news.jsp?id=65000417
>>
>>I find it hard to believe that executable format and execution model wouldn't affect performance somewhat. :)
>
>I don't believe I claimed otherwise. I simply said that effort was made in the construction of each
>SPEC CPU suite to minimize the effect of OS/library related cross platform differences. But there
>are OS functions that cannot be avoided, like virtual memory management. Perhaps the VM code
>in Windows is more efficient on the target machine than Linux's.
>