Hi Folks, I was wondering if anyone has an pointers/suggestions on how I might increase networking performance of a HVM NV DomU? Some simple benchmarks with iperf and nutcp show a decrease from about 938 Mbits/sec in the Dom0 to another machine down to only around 43.9 Mbits/sec in the DomU. (The Dom0 is NV) I really haven''t found much info on ways to tweak the network interface, does anyone have any pointers? Thanks... --joe
On Wed, Aug 08, 2007 at 04:16:56PM -0700, Joseph Mocker wrote:> I was wondering if anyone has an pointers/suggestions on how I might > increase networking performance of a HVM NV DomU?With an HVM guest using IO emulation we see much the same numbers as those you quote. There''s a suggestion that using an emulated AMD PCnet rather than RTL8139 will produce better results, but I haven''t tried that recently (last time I did the Solaris pcnet driver didn''t work with the qemu-dm emulation). We are working to provide PV IO drivers for Solaris when running under HVM, but they are not done yet. They should improve performance significantly. Is there a particular reason that you''re running NV with HVM rather than PV? Finally, if you are actually running a PV domU then the numbers are lower that those we''ve measured, though there is a lot of variance that we''re working to understand. dme.
David Edmondson wrote:> We are working to provide PV IO drivers for Solaris when running under > HVM, but they are not done yet. They should improve performance > significantly. >Excellent, I did read about PV drivers for Linux, so I was hoping the same was in the works for Solaris.> Is there a particular reason that you''re running NV with HVM rather > than PV? >I''ve just been trying to investigate different configurations with Xen, both PV and HVM. I did notice the PV numbers were higher. I guess I had also assumed that HVM would generally just perform better because "more" is done in hardware... which shows my lack of understanding since I''m not really sure what "more" is.> Finally, if you are actually running a PV domU then the numbers are > lower that those we''ve measured, though there is a lot of variance > that we''re working to understand. >Yes. I did notice with a PV Solaris DomU that performance was better. Thanks. --joe
On Thu, Aug 09, 2007 at 11:44:18AM -0700, Joseph Mocker wrote:> I''ve just been trying to investigate different configurations with Xen, > both PV and HVM. I did notice the PV numbers were higher. I guess I had > also assumed that HVM would generally just perform better because "more" is > done in hardware... which shows my lack of understanding since I''m not > really sure what "more" is.It''s almost always the opposite - PV performs better than HVM, at least today. Of course, using PV has other requirements (the OS has to be ported). dme.
David Edmondson wrote:> On Thu, Aug 09, 2007 at 11:44:18AM -0700, Joseph Mocker wrote: >> I''ve just been trying to investigate different configurations with Xen, >> both PV and HVM. I did notice the PV numbers were higher. I guess I had >> also assumed that HVM would generally just perform better because "more" is >> done in hardware... which shows my lack of understanding since I''m not >> really sure what "more" is. >The x86 hardware virt support is still rather poor from a performance standpoint. That should drastically improve with time, as both AMD and Intel have major improvements coming in future chips. I can''t really say more on this list, as much of that information is under NDA.> It''s almost always the opposite - PV performs better than HVM, at > least today. Of course, using PV has other requirements (the OS has to > be ported). > > dme.