Xen users - Nov 2010 - Unexpected memory bandwidth overhead with Xen

Hello Everyone,

While benchmarking a Dell Node (Dual Socket Quad Core AMD Opteron 2354 with 8 GB
of memory & CentOS 5.4 installed) , I am getting approx 25-30% overhead with
xen after running the STREAM memory bandwidth benchmark using 8 threads with
different compilers (gcc, pgi and icc).

These are the results with and without running xen kernel.
Centos:
Kernel Version: 2.6.18-194.17.1.el5
Kernel Version: 2.6.18-194.17.1.el5xen



 
 
 
 
 
  Function
  pgcc
  pgcc-xen
  Variation (%)
 
 
  Copy
  17108.0186
  11827.4346        

  30.86613432
 
 
  Scale
  16179.1128
  11692.2545
  27.7324125
 
 
  Add
  16706.097
  12212.2696
  26.89932544
 
 
  Triad
  17211.552
  12936.0666
  24.84078949
 
 
  

  

  

  

 
 
  Function
  icc
  icc-xen
  Variation (%)
 
 
  Copy
  16731.7459
  11743.8163
  29.81117231
 
 
  Scale
  16026.4231
  11560.9261
  27.86334151
 
 
  Add
  16656.4325
  12077.4699
  27.49065624
 
 
  Triad
  16701.5193
  12021.5076
  28.02147287
 
 
  

  

  

  

 
 
  Function
  gcc
  gcc-xen
  Variation (%)
 
 
  Copy
  11762.6266
  8810.7558
  25.09533712
 
 
  Scale
  11499.5329
  8611.5352
  25.11404355
 
 
  Add
  12399.1248
  9388.1446
  24.28381235
 
 
  Triad
  12607.9727
  9531.4749
  24.40120924
 
I know there are new kernels available but I never heard of such a bad
performance for bandwidth numbers for Xen. Does anyone of you know the reason?

Also, HPL on the other hand with the same setup gave a reasonably expected
performance overhead for Gflops count. It was approx 3.9-4% overhead with xen.

Does Xen behave like this with memory intensive applications?

Any comments and suggestions will really help.

Thanks,
Rohan




      

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