14 Replies Latest reply on Jul 8, 2008 3:09 PM by jmattson

    VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities

    prisoner881 Enthusiast

       

      I just upgraded my Beta 2 install to RC1 today and noticed a new error in the web GUI event log.  Upon starting my VM, it says hardware virtualization is not supported and that software virtualization will be used instead.  The CPU is an Athlon64 X2 6000+ which does support hardware virtualization.  The motherboard is a Tyan S3950 Socket AM2, and virtualization extensions are enabled in the BIOS.  Host OS is Windows Server 2008 Standard x64.  Guests are Windows Server 2008 Standard x64 and Windows Server 2003 R2 x64.

       

       

      Now I suppose it's possible RC1 is only considering Opteron CPU's as capable of hardware virtualization, but that's not true.  Please fix this.

       

       

        • 1. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
          SUJO Novice

          I have perhaps the same trouble or just similar one. I have noticed in event log (WebUI) that the "hardware MMUs are not supported by this host so using software instead". I have Intel Quad Core Xeon model E5405 which also does support hardware virtualization and all other kinds of implementations that adds to the virtualization techniques... I am not pointing blame yet but - I also am running most advanced system that also supports all extensions of CPU (Windows Server 2008) and all features are enabled in BIOS..

           

          I think that VMW should look into this...

           

           

          EDIT: same reports for EPT functions - "supposedly" they are not supported by my system...

           

           

           

           

           

          I am the First One.

          • 2. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
            rhsoftware Hot Shot

            I had to shut down the host completly after installting beta2, then VT worked and i was able to install 64Bit-Guets

            The WebUI said this also in the warning

            Just try shut down, take away the power cable. count until five and start the machine again

            • 3. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
              prisoner881 Enthusiast

               

              Sorry, the power down and wait thing didn't work for me.  And the problem is worse than I thought.

               

               

               

               

               

              I've got two hosts.  One is the Athlon64 X2 6000+ on a Tyan S3950, the other is a dual Opteron 2220 on a Tyan S3992.  Both of these machines fully support the hardware virtualization extensions.  Both have that feature turned on in the BIOS.  However, RC1 says neither of them support hardware virtualization.

               

               

               

               

               

              Now, I could see there being a bug with the Athlon64 X2 6000+ setup being improperly detected, but there is no excuse for the Opteron setup to suffer the same fate.  Clearly there's a problem, although I can hope it's nothing more than a logging bug (i.e. incorrectly reporting the use of software virtualization).

               

               

              • 4. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                Champion

                Due to hardware implementation bugs, we do not support hardware virtualization on AMD Family 0FH processors.  (AMD and not Intel, as previously stated).

                 

                Message was edited by: jmattson

                 

                Corrected processor vendor.  Thanks, Peter.

                • 5. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                  Champion

                  Intel has not yet shipped processors that support EPT.  The first such processor is Nehalem.  EPT is a second generation hardware virtualization feature.

                  • 6. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                    Peter_vm Guru
                    jmattson wrote:

                    Due to hardware implementation bugs, we do not support hardware virtualization on Intel Family 0FH processors.

                    I believe OP referenced only AMD processors, not Intel?

                    • 7. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                      prisoner881 Enthusiast

                       

                      +"Due to hardware implementation bugs, we do not support hardware

                      virtualization on AMD Family 0FH processors. (AMD and not Intel, as

                      previously stated)."+

                       

                       

                       

                       

                       

                      OK, so the Athlon64 X2 6000+ hardware virtualization isn't supported.  I can accept that.  Why then do I get the same message on my Opteron 2220?  Or are you saying that every AMD CPU that isn't quad core lacks support?  

                       

                       

                       

                       

                       

                      If it's the latter then I will be very disappointed.  I find it difficult to believe that VMWare, the company that leads the world in virtualization technology, can't use AMD-V on these chips.  It almost sounds like VMWare had decided not to pursue it vigorously enough instead of it being an "implementation bug" as you suggest.  The fact that Intel bought a $200 million stake in VMWare last year also make me wonder if VMWare is playing favorites.  Intel has a seat on the VMWare board, and I'm sure they're not too unhappy that VMWare won't support AMD-V on the most popular chips AMD sells. This is the first I've ever heard of someone calling AMD-V's

                      implementation buggy.  I can find no details of it anywhere on the web.

                       

                       

                       

                       

                       

                      I wonder if Hyper-V supports AMD-V on the 0FH line?  Wouldn't it be funny if the guys at Microsoft have figured out a way to do something that the vaunted VMWare programmers could not -- or is it would not -- do?

                       

                       

                      • 8. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                        Champion
                        prisoner881 wrote:

                        OK, so the Athlon64 X2 6000+ hardware virtualization isn't supported.  I can accept that.  Why then do I get the same message on my Opteron 2220?  Or are you saying that every AMD CPU that isn't quad core lacks support?  

                         

                        To use AMD-V, we require a bugfix that was made in AMD Family 10H processors.  The software workaround for the Family 0FH chip bug (or "feature", if you prefer) was deemed to be too complex, not entirely correct, and an unnecessary drain on QA resources.  This decision was also supported by results from our performance group which indicated that our binary translation implementation outperformed our AMD-V implementation on Family 0FH processors.

                         

                        The best virtualization implementation that we have ever developed for Family 0FH processors is the one that uses binary translation.  You can blame that on politics, or on incompetence, or on an additional 10 years of development.  Take your pick.  In any case, you really are not missing anything, unless you want to run nested VMs.  That is a niche application which we did not consider when making our decision not to support AMD-V on Family 0FH.

                        • 9. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                          prisoner881 Enthusiast
                          jmattson wrote:
                          prisoner881 wrote:

                          OK, so the Athlon64 X2 6000+ hardware virtualization isn't supported.  I can accept that.  Why then do I get the same message on my Opteron 2220?  Or are you saying that every AMD CPU that isn't quad core lacks support?  

                          To use AMD-V, we require a bugfix that was made in AMD Family 10H processors.  The software workaround for the Family 0FH chip bug (or "feature", if you prefer) was deemed to be too complex, not entirely correct, and an unnecessary drain on QA resources.  This decision was also supported by results from our performance group which indicated that our binary translation implementation outperformed our AMD-V implementation on Family 0FH processors.

                           

                          The best virtualization implementation that we have ever developed for Family 0FH processors is the one that uses binary translation.  You can blame that on politics, or on incompetence, or on an additional 10 years of development.  Take your pick.  In any case, you really are not missing anything, unless you want to run nested VMs.  That is a niche application which we did not consider when making our decision not to support AMD-V on Family 0FH.

                           

                           

                           

                          Then what advantage (if any) does the "fixed" AMD-V extensions give on the 23xx Opterons?  Not the nested paging thing, which I agree is a rare circumstance.  Is there any advantage to hardware virtualization outside of nested VM's?  What would be the advantage of such a thing in the first place?

                           

                           

                           

                           

                           

                          And not to cast doubt on what you say, but I simply cannot find anyone on the Xen forums complaining about broken implementations with 0FH family CPU's. Hyper-V is too new to have much data on anything at the moment so that's out as well.  I'd be fascinated to know what the actual fault was with 0FH and how it was fixed in the latest gen CPU's.

                           

                           

                           

                           

                           

                          Last but certainly not least, does ESX suffer from the same problem, or is that a different can of worms?

                           

                           

                          • 10. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                            Champion
                            prisoner881 wrote:

                            Then what advantage (if any) does the "fixed" AMD-V extensions give on the 23xx Opterons?  Not the nested paging thing, which I agree is a rare circumstance.  Is there any advantage to hardware virtualization outside of nested VM's?  What would be the advantage of such a thing in the first place?

                             

                            Let's not confuse nested paging with nested VMs.  Nested paging (a.k.a. "rapid virtualization indexing") offers hardware virtualization of the MMU.  When we talk about nested VMs, we're talking about running one virtual machine inside another.  Nested paging eliminates 70% - 90% of the #vmexits from the guest to the hypervisor for typical applications.  By eliminating the major source of hypervisor intercepts, nested paging offers a dramatic increase in hardware virtualization performance.  Our implementation of AMD-V with RVI generally outperforms our binary translation implementation on the same hardware.

                             

                            And not to cast doubt on what you say, but I simply cannot find anyone on the Xen forums complaining about broken implementations with 0FH family CPU's. Hyper-V is too new to have much data on anything at the moment so that's out as well.  I'd be fascinated to know what the actual fault was with 0FH and how it was fixed in the latest gen CPU's.

                             

                            Family 0FH processors fail to populate the EXITINTINFO field of the VMCB when an IDT vectoring event is interrupted by a task switch intercept.  Family 10H processors properly populate the field with information about the interrupted IDT vectoring event.

                             

                            Last but certainly not least, does ESX suffer from the same problem, or is that a different can of worms?

                             

                            ESX 3.5 also supports AMD-V only on Family 10H processors.

                            • 11. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                              prisoner881 Enthusiast

                               

                              Thank you for the detailed information on the issue.  Now for a really loaded question: which CPU benefits more from hardware virtualization, Intel or AMD?  What I'm really trying to get a handle on is which way I should go:  replace my older 22xx Opterons with 23xx Opterons (more cores but slower clocks) or ditch the servers entirely and go with Intel?  Note that VM performance thus far seems perfectly acceptable, but load is undoubtedly going to grow over time. If hardware virtualization is really worth it, though, it might warrant earlier replacement of these machines.

                               

                               

                               

                               

                               

                              Further, I can't help but feel AMD is in a better position to capitalize on the VM movement, what with their HyperTransport architecture and the far better performance when you go above 4 CPU's. Until Intel gets rid of the FSB's and implements an on-chip memory controller, it's always going to have less available CPU bandwidth in machines with lots of sockets.  And lots of sockets is exactly what VM's like...along with lots of RAM and fast drive arrays, of course.

                               

                               

                              • 12. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                                Champion

                                At present, the CPU vendors are in a very competitive position with respect to hardware virtualization.  With hardware virtualization, both AMD Family 10H processors and Intel 45nm processors achieve better performance than with binary translation.  With AMD Family 10H processors, a significant boost is due to Rapid Virtualization Indexing (or nested paging).  With Intel 45nm processors, the boost is mainly due to improved latencies for VM-entry and VM-exit.  Also, Intel's FlexPriority, introduced with Clovertown, offers a boost for certain 32-bit Windows guests (Windows 2000 and Windows XP, in particular).  The Achilles' Heel of Rapid Virtualization Indexing is the added cost of a nested page walk to satisfy a TLB miss.  Without hardware virtualization of the MMU, Intel's chips suffer on workloads with high MMU activity.  Also, as you point out, there may be platform issues to consider in addition to CPU characteristics.

                                1 person found this helpful
                                • 13. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                                  prisoner881 Enthusiast

                                  You say the latest generation of both CPU's offers performance above and beyond that of binary translation.  Can you give me a ballpark percentage performance increase figure compared on a clock-for-clock basis with binary translation?  I know it's probably something that varies wildly based on the workload, but surely you have some general rule-of-thumb that you've seen in your experiences.  Are we talking a 10% speedup or a 500% speedup or something in between?

                                  • 14. Re: VMWare Server 2 RC1 incorrectly ID's hardware virtualization capabilities
                                    Champion

                                    There is considerable variation, but as a rule of thumb, you can expect hardware virtualization with current processors to be anywhere from 5% slower to 15% faster than binary translation, depending on the application.