[Xen-devel] unusual domU memory performance

["Bryan D. Payne" <bryan@xxxxxxxxxxxx>]

I've been experiencing some unexpected results with regards to the  
memory performance of a domU virtual machine.  I'm hoping that  
someone with a little more familiarity with Xen can shed some light  
on why I might be seeing this behavior.  The scenario is as follows...
I have a program that runs in dom0 and maps some kernel memory from  
domU (the memory that I'm mapping holds the system call table).  Once  
the memory is mapped, I run in a loop and compare the table against a  
known good value for the table.  I have the loop setup so that I can  
adjust how frequently the memory comparison is made.  The weird thing  
is that when I make the comparison continuously (i.e., no sleep  
between comparisons), the memory performance of domU actually  
increases compared to a baseline without my memory checker running at  
all.  In fact, using a sleep of 0.5sec or 1.0sec still creates a  
faster memory benchmark in domU.  Once I get to a sleep interval of  
1.5sec or greater, the memory benchmark is finally slower than the  
baseline.
Some more details for those that are interested:
* You can view the data in graph form at http://www.cc.gatech.edu/ 
~bdpayne/notes/copy.pdf.
* I'm using the stream memory benchmark.  The data that I show in the  
graph above is for copy, but similar trends are seen for scale, add,  
and triad.
* Each point in the graph is an average of 3000 benchmark tests,  
taken in 3 sets of 1000 tests where each test in the 1000 was  
separated by 2 seconds.
* The bars in the graph indicate the statistical error interval for  
each set of 3000 benchmarks
* Note that there is more variability in the measurements from the  
continuous test than the tests with some sleep interval inserted in  
the loop (wider error interval)
These results were somewhat counter-intuitive to me as I didn't  
expect memory performance to improve while I was doing extra  
monitoring.  Does anyone have some thoughts on why I'd be seeing this  
behavior?  I'm happy to provide additional data if that's helpful.
Thanks,
bryan

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel