Hi Keith, Joost,

Arriving late to this thread, I can definitely lend my support that autoconf is *painfully* slow on a virtualized windows instance under VMWare, under either Mac OS X 10.5 and Ubuntu 9.04 hosts.

Keith, we ofter hear that (to summarize) "it's not mingw's fault, it's Windows which is intrinsically slow".  For my own understanding, I would love to hear a detailed explanation of what you think might be going on.  Is it that Windows is really bad at opening file descriptors?  Starting or switching processes? etc.  (And if you've already discussed this on another thread, feel free to simply point me there, too.)

Keith, you recommend a linux-hosted cross tool chain for building Windows apps.  On mingw.org, I see the page " http://mingw.org/wiki/LinuxCrossMinGW", but that page hasn't been updated since Sept 2008 and I know (from watching this list) that there's been a lot of updates since.  Could you please point me to more recent documentation, if any exists?

Many thanks to Keith and the mingw team for their work on this very useful project.

-Natalie


On 3/4/10 3:20 AM, Keith Marshall wrote:
On Wednesday 03 March 2010 12:03:44 Joost Kraaijeveld wrote:
  
What I see is the user-part is more or less the same in all three
experiments but it is the system part that takes very long in the
VirtualBox situation. This is the same in Keith's output (I assume
he configured the same project three times).
    
Yes, I ran configure on the mingw-get source snapshot I posted a week 
or so ago, in all three cases.  However, you seem to be jumping to a 
conclusion which I *don't* think is *significantly* apparent in my 
data, (which seems reasonably representative and reproducible over 
three or four repetitions for each of the three cases, BTW).  Yes, I 
can see that system time is 48.3% of elapsed time for the VirtualBox 
case, compared to 42.5% for the real Win2K case, but I can't say if 
this is a statistically significant difference, because I haven't 
performed enough tests; however, the difference between the two MSYS 
cases and the native Linux case at 17.2% likely is significant.

A similar subjective analysis of the user mode times shows 56.1% for 
the real Win2K case, 25.6% for the VirtualBox case and 54.9% for the 
native Linux case.  Once again, there may be no significance, but if 
anything my VirtualBox marginally outperforms the real Win2K host, 
even when run on less highly specified hardware.

The only conclusion *I* will draw from this is that it is a waste of 
time setting up a virtual MS-Windows guest on a Linux host, to run 
MSYS and MinGW for building MS-Windows applications; rather set up a 
cross tool chain, directly hosted on the Linux box, and benefit from 
a performance boost of around 1000% .. 1400%, in terms of elapsed 
time, over any MS-Windows hosted build.