Re: [User-mode-linux-user] lmbench numbers?

[Jeff D. <jd...@ka...>]

> User-mode Linux seems to be about 100 times slower than Linux on bare
> hardware on fast operations.  Slower operations like fork() and exec()
> does better. 

I was expecting it to be pretty bad.  Especially things like system calls.  On 
bare hardware a system call is not too many instructions.  On uml, it's a 
whole lot of context switches.

> Surprisingly, user-mode Linux seems to be do mmap() and page faults
> faster than Linux on hardware.  I wonder why? 

Maybe the hosting kernel usually has the data already in memory, so it's just 
a matter of mapping it in, whereas on bare hardware, the data has to be read 
in more often from disk?

> The benchmark crashed the kernel before finishing the context switch
> and memory latency measurements.

Looks like I've got some debugging to do.  Where do you get lmbench?

> When I try to run the context switch benchmark, uml forks off lots of
> subprocesses and I get a lot of
>         fork: Resource temporarily unavailable messages from the
> hosting kernel. 

What's lmbench doing at that point?  Is it possible that uml just isn't 
killing off dead processes quickly enough?

				Jeff