Hi, I''ve made some SAP benchmarks on my Xen system and discovered a huge difference in the performance of a "xened" SAP system compared to a native SAP system. Hence, I tried to figure out what might cause this ''overhead'' and run a xentrace (listining to all events). Xentrace produced 24gb data and I converted it to 27gb human-readable data by using xentrace_format. After I gathered the human-readable data, I filtered the data and counted the appearance of each event. So far, so good. Now it comes: although I used paravirt-guests, the xentrace-tool reported HVM events in the trace data. Moreover, from my point of view it is impossible to trace HVM events, as I use AMD Opteron 280 with no AMD-V feature. Did I miss something or does the xentrace-tool report ''wrong'' trace data? I use Xen version 3.2.0_16718_14-0.4 (from Suse). Thanks André _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Are you using xentrace_format, xentrace, and xen all from the same version ? I don''t think the trace-ids are incompatible across versions, but still. Also, what HVM events are you seeing ? -dulloor On Mon, May 17, 2010 at 11:58 PM, André Bögelsack <Andre.Boegelsack@in.tum.de> wrote:> Hi, > > I''ve made some SAP benchmarks on my Xen system and discovered a huge > difference in the performance of a "xened" SAP system compared to a native > SAP system. Hence, I tried to figure out what might cause this ''overhead'' > and run a xentrace (listining to all events). Xentrace produced 24gb data > and I converted it to 27gb human-readable data by using xentrace_format. > After I gathered the human-readable data, I filtered the data and counted > the appearance of each event. So far, so good. > > Now it comes: although I used paravirt-guests, the xentrace-tool reported > HVM events in the trace data. Moreover, from my point of view it is > impossible to trace HVM events, as I use AMD Opteron 280 with no AMD-V > feature. > > Did I miss something or does the xentrace-tool report ''wrong'' trace data? > > I use Xen version 3.2.0_16718_14-0.4 (from Suse). > > Thanks > André > _______________________________________________ > Xen-devel mailing list > Xen-devel@lists.xensource.com > http://lists.xensource.com/xen-devel > >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
I didn''t change anything so I assume it is the same version across the tools. These are the events I gathered from xentrace: It is not possible to call a HVM event on that hardware platform - the CPU is too old. My guess is, that the trace_id''s recorded by xentrace and interpreted by xentrace_format are not identical. Is there any way to prove it? Regards André Dulloor <dulloor@gmail.com> schrieb am 03.06.2010 22:02:02:> Von: Dulloor <dulloor@gmail.com> > An: André Bögelsack <Andre.Boegelsack@in.tum.de> > Kopie: xen-devel@lists.xensource.com > Datum: 03.06.2010 22:02 > Betreff: Re: [Xen-devel] Question regarding xentrace > > Are you using xentrace_format, xentrace, and xen all from the same > version ? I don''t think the trace-ids are incompatible across > versions, but still. > Also, what HVM events are you seeing ? > > -dulloor > > On Mon, May 17, 2010 at 11:58 PM, André Bögelsack > <Andre.Boegelsack@in.tum.de> wrote: > > Hi, > > > > I''ve made some SAP benchmarks on my Xen system and discovered a huge > > difference in the performance of a "xened" SAP system compared to anative> > SAP system. Hence, I tried to figure out what might cause this''overhead''> > and run a xentrace (listining to all events). Xentrace produced 24gbdata> > and I converted it to 27gb human-readable data by usingxentrace_format.> > After I gathered the human-readable data, I filtered the data andcounted> > the appearance of each event. So far, so good. > > > > Now it comes: although I used paravirt-guests, the xentrace-toolreported> > HVM events in the trace data. Moreover, from my point of view it is > > impossible to trace HVM events, as I use AMD Opteron 280 with no AMD-V > > feature. > > > > Did I miss something or does the xentrace-tool report ''wrong'' tracedata?> > > > I use Xen version 3.2.0_16718_14-0.4 (from Suse). > > > > Thanks > > André > > _______________________________________________ > > Xen-devel mailing list > > Xen-devel@lists.xensource.com > > http://lists.xensource.com/xen-devel > > > >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel
Hmm, that would be rather exceptional... the trace_id''s haven''t really changed since tracing was introduced, only new ones added. Are you sure you''re using the xentrace_format that matches up to the 3.2 hypervisor you''re using (and not, say, something from -unstable)? The format changed from fixed-length records to variable-length records somewhere around there (can''t remember exactly). If you take a really short trace (i.e., just let it run for 5 seconds before interrupting), I can try to take a quick look at it, to see what I see. -George On Mon, Jun 7, 2010 at 2:50 PM, André Bögelsack <Andre.Boegelsack@in.tum.de>wrote:> I didn''t change anything so I assume it is the same version across the > tools. > > These are the events I gathered from xentrace: > > > It is not possible to call a HVM event on that hardware platform - the CPU > is too old. > > My guess is, that the trace_id''s recorded by xentrace and interpreted by > xentrace_format are not identical. Is there any way to prove it? > > Regards > André > > Dulloor <dulloor@gmail.com> schrieb am 03.06.2010 22:02:02: > > > Von: Dulloor <dulloor@gmail.com> > > An: André Bögelsack <Andre.Boegelsack@in.tum.de> > > Kopie: xen-devel@lists.xensource.com > > Datum: 03.06.2010 22:02 > > Betreff: Re: [Xen-devel] Question regarding xentrace > > > > Are you using xentrace_format, xentrace, and xen all from the same > > version ? I don''t think the trace-ids are incompatible across > > versions, but still. > > Also, what HVM events are you seeing ? > > > > -dulloor > > > > On Mon, May 17, 2010 at 11:58 PM, André Bögelsack > > <Andre.Boegelsack@in.tum.de> wrote: > > > Hi, > > > > > > I''ve made some SAP benchmarks on my Xen system and discovered a huge > > > difference in the performance of a "xened" SAP system compared to a > native > > > SAP system. Hence, I tried to figure out what might cause this > ''overhead'' > > > and run a xentrace (listining to all events). Xentrace produced 24gb > data > > > and I converted it to 27gb human-readable data by using > xentrace_format. > > > After I gathered the human-readable data, I filtered the data and > counted > > > the appearance of each event. So far, so good. > > > > > > Now it comes: although I used paravirt-guests, the xentrace-tool > reported > > > HVM events in the trace data. Moreover, from my point of view it is > > > impossible to trace HVM events, as I use AMD Opteron 280 with no AMD-V > > > feature. > > > > > > Did I miss something or does the xentrace-tool report ''wrong'' trace > data? > > > > > > I use Xen version 3.2.0_16718_14-0.4 (from Suse). > > > > > > Thanks > > > André > > > _______________________________________________ > > > Xen-devel mailing list > > > Xen-devel@lists.xensource.com > > > http://lists.xensource.com/xen-devel > > > > > > > > _______________________________________________ > Xen-devel mailing list > Xen-devel@lists.xensource.com > http://lists.xensource.com/xen-devel > >_______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel