CPU sockets, threads, cores and NUMA
Hello developers, tl;dr version: * deprecate report_host_threads_as_cores * remove cpuSockets, use sum(numaNodes.keys()) * report threadsPerCore for ppc64le / report total number of threads for ppc64le * work on our naming issues I've been going over our capabilities reporting code in VDSM due to specific threading requirements on ppc64le platform and noticed few issues. Before trying to fix something that "works", I'm sending this mail to start a discussion regarding current and future state of the code. First thing is the terminology. What we consider cpu sockets, cores and threads are in fact NUMA cells, sum of cores present in NUMA nodes and the same for threads. I'd like to see the code moving in a direction that is correct in this sense. More important are the actual calculations. I believe we should draw an uncrossable line between cores and threads and not interfere with it at least on VDSM's side. That would mean deprecating report_host_threads_as_cores option. The actual algorithm used at present does calculate the numa cores and numa threads correctly given that there are no offline CPUs - most likely fine enough. We don't have to report the actual number of sockets though, as it is reported in numa* keys. It does fail to provide us with information that can be used in ppc64le environment, where for POWER8 we want to run the host without SMT while VMs would have multiple CPUs assigned. There are various configurations of so-called subcores in POWER8, where each CPU core can contain 1, 2 or 4 subcores. This configuration must be taken in consideration as given e.g. 160 threads overall, it is possible to run either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2 mode. We have to report either the total number of threads OR just the threadsPerCore setting, so the users know how many "CPUs" should be assigned to machines for optimal performance. As always, I welcome any opinions regarding the proposed ideas. Also note that all of the changes can be done via deprecation to be fully backwards compatible - except for the ppc part. Regards, mpolednik
----- Original Message -----
From: "Martin Polednik" <mpolednik@redhat.com> To: devel@ovirt.org Sent: Thursday, December 10, 2015 4:07:57 PM Subject: [ovirt-devel] CPU sockets, threads, cores and NUMA
It does fail to provide us with information that can be used in ppc64le environment, where for POWER8 we want to run the host without SMT while VMs would have multiple CPUs assigned. There are various configurations of so-called subcores in POWER8, where each CPU core can contain 1, 2 or 4 subcores.
What are those subcores? Does this subcore concept map on something like SMT on x86_64? Or is this an hardware concept (mind goes to the AMD's bulldozer modules)
This configuration must be taken in consideration as given e.g. 160 threads overall, it is possible to run either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2 mode.
Could you elaborate a bit more about the se smtX modes and how they play with threads and cores?
We have to report either the total number of threads OR just the threadsPerCore setting, so the users know how many "CPUs" should be assigned to machines for optimal performance.
Which options is best for ppc, if there is any difference? bests, -- Francesco Romani RedHat Engineering Virtualization R & D Phone: 8261328 IRC: fromani
On Thu, Dec 10, 2015 at 5:07 PM, Martin Polednik <mpolednik@redhat.com> wrote:
Hello developers,
tl;dr version: * deprecate report_host_threads_as_cores * remove cpuSockets, use sum(numaNodes.keys()) * report threadsPerCore for ppc64le / report total number of threads for ppc64le * work on our naming issues
I've been going over our capabilities reporting code in VDSM due to specific threading requirements on ppc64le platform and noticed few issues. Before trying to fix something that "works", I'm sending this mail to start a discussion regarding current and future state of the code.
First thing is the terminology. What we consider cpu sockets, cores and threads are in fact NUMA cells, sum of cores present in NUMA nodes and the same for threads. I'd like to see the code moving in a direction that is correct in this sense.
Note that I think users are more familiar with sockets-cores-threads than NUMA cells, terminology-wise.
More important are the actual calculations. I believe we should draw an uncrossable line between cores and threads and not interfere with it at least on VDSM's side. That would mean deprecating report_host_threads_as_cores option. The actual algorithm used at present does calculate the numa cores and numa threads correctly given that there are no offline CPUs - most likely fine enough. We don't have to report the actual number of sockets though, as it is reported in numa* keys.
There is a reason for report_host_threads_as_cores option. I don't remember it right now, but it had to do with some limitation of some OS or license or something. I don't think we should deprecate it.
It does fail to provide us with information that can be used in ppc64le environment, where for POWER8 we want to run the host without SMT while VMs would have multiple CPUs assigned. There are various configurations of so-called subcores in POWER8, where each CPU core can contain 1, 2 or 4 subcores. This configuration must be taken in consideration as given e.g. 160 threads overall, it is possible to run either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2 mode. We have to report either the total number of threads OR just the threadsPerCore setting, so the users know how many "CPUs" should be assigned to machines for optimal performance.
YAY... do we have a comparison what libvirt knows / looks at (or they ignore it altogether?) Y.
As always, I welcome any opinions regarding the proposed ideas. Also note that all of the changes can be done via deprecation to be fully backwards compatible - except for the ppc part.
Regards, mpolednik _______________________________________________ Devel mailing list Devel@ovirt.org http://lists.ovirt.org/mailman/listinfo/devel
On 10 Dec 2015, at 16:36, Yaniv Kaul <ykaul@redhat.com> wrote: =20 On Thu, Dec 10, 2015 at 5:07 PM, Martin Polednik <mpolednik@redhat.com = <mailto:mpolednik@redhat.com>> wrote: Hello developers, =20 tl;dr version: * deprecate report_host_threads_as_cores * remove cpuSockets, use sum(numaNodes.keys()) * report threadsPerCore for ppc64le / report total number of threads for ppc64le * work on our naming issues =20 I've been going over our capabilities reporting code in VDSM due to specific threading requirements on ppc64le platform and noticed few issues. Before trying to fix something that "works", I'm sending this mail to start a discussion regarding current and future state of the code. =20 First thing is the terminology. What we consider cpu sockets, cores = and threads are in fact NUMA cells, sum of cores present in NUMA nodes and the same for threads. I'd like to see the code moving in a direction that is correct in this sense. =20 Note that I think users are more familiar with sockets-cores-threads =
--Apple-Mail=_643EEE38-B765-4581-8070-B2E414E4F79B Content-Transfer-Encoding: quoted-printable Content-Type: text/plain; charset=utf-8 than NUMA cells, terminology-wise. we do report numa separately today, and we should keep doing that. I = consider it another level of detail/complexity which many users do not = care about.=20 So we should keep both
=20 =20 More important are the actual calculations. I believe we should draw an uncrossable line between cores and threads and not interfere with it at least on VDSM's side. That would mean deprecating report_host_threads_as_cores option. The actual algorithm used at present does calculate the numa cores and numa threads correctly given that there are no offline CPUs - most likely fine enough. We don't have to report the actual number of sockets though, as it is reported in numa* keys. =20 There is a reason for report_host_threads_as_cores option. I don't = remember it right now, but it had to do with some limitation of some OS = or license or something. I don't think we should deprecate it.
the idea was to remove that option from VDSM conf (as it=E2=80=99s = cumbersome to use), and rather report all relevant information so engine = can decide later on whether to count it this way or another Today it=E2=80=99s used as a simple =E2=80=9Ccore multiplier=E2=80=9D if = your workload is running "good enough=E2=80=9D in parallel on 2 threads = within one core, we just consider it as additional available =E2=80=9Ccpu=E2= =80=9D. For some workloads where this assumption is not working well, = and also for licensing or any other reason, you can disable it and see = =E2=80=9Chalf=E2=80=9D of the cpus on x86 despite having HT enabled in = BIOS. On PPC this is more tricky as MArtin says below - (threads are not able = to run multiple VMs simultaneously) - so we need to push that decision = from vdsm up the chain.
=20 It does fail to provide us with information that can be used in ppc64le environment, where for POWER8 we want to run the host without SMT while VMs would have multiple CPUs assigned. There are various configurations of so-called subcores in POWER8, where each CPU core can contain 1, 2 or 4 subcores. This configuration must be taken in consideration as given e.g. 160 threads overall, it is possible to run either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2 mode. We have to report either the total number of threads OR just the threadsPerCore setting, so the users know how many "CPUs" should be assigned to machines for optimal performance.
x per y sounds best to me but I think it=E2=80=99s even more complicated, if we consider offline = CPUs(we don=E2=80=99t do that today) then the default picture on POWER8 = currently looks like 20 cores in 4 numa cells, 8 threads per core. SMT = is disabled altogether, so CPUs 1-7,9-15,=E2=80=A6 are offline. So = should we report them or not? On x86 I would not do that as they are = administratively disabled and can=E2=80=99t be used, however on ppc = since RHEL 7.2 they are dynamically enabled on demand (if the guest = topology uses threads as well), so they should be reported as available = (or "sort-of-available=E2=80=9D:) still, I think we should go with simple "sockets, cores/socket, = threads/core=E2=80=9D numbers, the rest needs to be computed or chosen from, based on additional = detailed report of NUMA topology and online/offline CPU status perhaps with different behavior/capabilities on x86 and on power
=20 YAY... do we have a comparison what libvirt knows / looks at (or they = ignore it altogether?) Y. =20 =20 As always, I welcome any opinions regarding the proposed ideas. Also note that all of the changes can be done via deprecation to be fully backwards compatible - except for the ppc part. =20 Regards, mpolednik _______________________________________________ Devel mailing list Devel@ovirt.org <mailto:Devel@ovirt.org> http://lists.ovirt.org/mailman/listinfo/devel = <http://lists.ovirt.org/mailman/listinfo/devel> =20 _______________________________________________ Devel mailing list Devel@ovirt.org http://lists.ovirt.org/mailman/listinfo/devel
--Apple-Mail=_643EEE38-B765-4581-8070-B2E414E4F79B Content-Transfer-Encoding: quoted-printable Content-Type: text/html; charset=utf-8 <html><head><meta http-equiv=3D"Content-Type" content=3D"text/html = charset=3Dutf-8"></head><body style=3D"word-wrap: break-word; = -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" = class=3D""><br class=3D""><div><blockquote type=3D"cite" class=3D""><div = class=3D"">On 10 Dec 2015, at 16:36, Yaniv Kaul <<a = href=3D"mailto:ykaul@redhat.com" class=3D"">ykaul@redhat.com</a>> = wrote:</div><br class=3D"Apple-interchange-newline"><div class=3D""><div = dir=3D"ltr" class=3D""><div class=3D"gmail_extra"><div = class=3D"gmail_quote">On Thu, Dec 10, 2015 at 5:07 PM, Martin Polednik = <span dir=3D"ltr" class=3D""><<a href=3D"mailto:mpolednik@redhat.com" = target=3D"_blank" class=3D"">mpolednik@redhat.com</a>></span> = wrote:<br class=3D""><blockquote class=3D"gmail_quote" style=3D"margin:0 = 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">Hello = developers,<br class=3D""> <br class=3D""> tl;dr version:<br class=3D""> * deprecate report_host_threads_as_cores<br class=3D""> * remove cpuSockets, use sum(numaNodes.keys())<br class=3D""> * report threadsPerCore for ppc64le / report total number of threads<br = class=3D""> for ppc64le<br class=3D""> * work on our naming issues<br class=3D""> <br class=3D""> I've been going over our capabilities reporting code in VDSM due to<br = class=3D""> specific threading requirements on ppc64le platform and noticed few<br = class=3D""> issues. Before trying to fix something that "works", I'm sending this<br = class=3D""> mail to start a discussion regarding current and future state of the<br = class=3D""> code.<br class=3D""> <br class=3D""> First thing is the terminology. What we consider cpu sockets, cores and = threads are in fact NUMA cells, sum of cores present in NUMA<br = class=3D""> nodes and the same for threads. I'd like to see the code moving in a<br = class=3D""> direction that is correct in this sense.<br class=3D""></blockquote><div = class=3D""><br class=3D""></div><div class=3D"">Note that I think users = are more familiar with sockets-cores-threads than NUMA cells, = terminology-wise.</div></div></div></div></div></blockquote><div><br = class=3D""></div></div><div>we do report numa separately today, and we = should keep doing that. I consider it another level of detail/complexity = which many users do not care about. </div><div>So we should keep = both</div><div><br class=3D""><blockquote type=3D"cite" class=3D""><div = class=3D""><div dir=3D"ltr" class=3D""><div class=3D"gmail_extra"><div = class=3D"gmail_quote"><div class=3D""> </div><blockquote = class=3D"gmail_quote" style=3D"margin:0 0 0 .8ex;border-left:1px #ccc = solid;padding-left:1ex"> <br class=3D""> More important are the actual calculations. I believe we should draw<br = class=3D""> an uncrossable line between cores and threads and not interfere with<br = class=3D""> it at least on VDSM's side. That would mean deprecating<br class=3D""> report_host_threads_as_cores option. The actual algorithm used at<br = class=3D""> present does calculate the numa cores and numa threads correctly = given<br class=3D""> that there are no offline CPUs - most likely fine enough. We don't<br = class=3D""> have to report the actual number of sockets though, as it is reported<br = class=3D""> in numa* keys.<br class=3D""></blockquote><div class=3D""><br = class=3D""></div><div class=3D"">There is a reason for = report_host_threads_as_cores option. I don't remember it right now, but = it had to do with some limitation of some OS or license or = something.</div><div class=3D"">I don't think we should deprecate = it.</div></div></div></div></div></blockquote><div><br = class=3D""></div><div>the idea was to remove that option from VDSM conf = (as it=E2=80=99s cumbersome to use), and rather report all relevant = information so engine can decide later on whether to count it this way = or another</div><div>Today it=E2=80=99s used as a simple =E2=80=9Ccore = multiplier=E2=80=9D if your workload is running "good enough=E2=80=9D in = parallel on 2 threads within one core, we just consider it as additional = available =E2=80=9Ccpu=E2=80=9D. For some workloads where this = assumption is not working well, and also for licensing or any other = reason, you can disable it and see =E2=80=9Chalf=E2=80=9D of the cpus on = x86 despite having HT enabled in BIOS.</div><div><br class=3D""></div>On = PPC this is more tricky as MArtin says below - (threads are not able to = run multiple VMs simultaneously) - so we need to push that decision from = vdsm up the chain.</div><div><br class=3D""><blockquote type=3D"cite" = class=3D""><div class=3D""><div dir=3D"ltr" class=3D""><div = class=3D"gmail_extra"><div class=3D"gmail_quote"><blockquote = class=3D"gmail_quote" style=3D"margin:0 0 0 .8ex;border-left:1px #ccc = solid;padding-left:1ex"> <br class=3D""> It does fail to provide us with information that can be used in<br = class=3D""> ppc64le environment, where for POWER8 we want to run the host without<br = class=3D""> SMT while VMs would have multiple CPUs assigned. There are various<br = class=3D""> configurations of so-called subcores in POWER8, where each CPU core<br = class=3D""> can contain 1, 2 or 4 subcores. This configuration must be taken in<br = class=3D""> consideration as given e.g. 160 threads overall, it is possible to = run<br class=3D""> either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2<br = class=3D""> mode. We have to report either the total number of threads OR just = the<br class=3D""> threadsPerCore setting, so the users know how many "CPUs" should be<br = class=3D""> assigned to machines for optimal performance.<br = class=3D""></blockquote></div></div></div></div></blockquote><div><br = class=3D""></div>x per y sounds best to me</div><div>but I think it=E2=80=99= s even more complicated, if we consider offline CPUs(we don=E2=80=99t do = that today) then the default picture on POWER8 currently looks like 20 = cores in 4 numa cells, 8 threads per core. SMT is disabled altogether, = so CPUs 1-7,9-15,=E2=80=A6 are offline. So should we report them or not? = On x86 I would not do that as they are administratively disabled and = can=E2=80=99t be used, however on ppc since RHEL 7.2 they are = dynamically enabled on demand (if the guest topology uses threads as = well), so they should be reported as available (or = "sort-of-available=E2=80=9D:)</div><div><br class=3D""></div><div>still, = I think we should go with simple "sockets, cores/socket, threads/core=E2=80= =9D numbers,</div><div>the rest needs to be computed or chosen from, = based on additional detailed report of NUMA topology and online/offline = CPU status</div><div>perhaps with different behavior/capabilities on x86 = and on power</div><div><br class=3D""><blockquote type=3D"cite" = class=3D""><div class=3D""><div dir=3D"ltr" class=3D""><div = class=3D"gmail_extra"><div class=3D"gmail_quote"><div class=3D""><br = class=3D""></div><div class=3D"">YAY... do we have a comparison what = libvirt knows / looks at (or they ignore it altogether?)</div><div = class=3D"">Y.</div><div class=3D""> </div><blockquote = class=3D"gmail_quote" style=3D"margin:0 0 0 .8ex;border-left:1px #ccc = solid;padding-left:1ex"> <br class=3D""> As always, I welcome any opinions regarding the proposed ideas. Also<br = class=3D""> note that all of the changes can be done via deprecation to be fully<br = class=3D""> backwards compatible - except for the ppc part.<br class=3D""> <br class=3D""> Regards,<br class=3D""> mpolednik<br class=3D""> _______________________________________________<br class=3D""> Devel mailing list<br class=3D""> <a href=3D"mailto:Devel@ovirt.org" target=3D"_blank" = class=3D"">Devel@ovirt.org</a><br class=3D""> <a href=3D"http://lists.ovirt.org/mailman/listinfo/devel" = rel=3D"noreferrer" target=3D"_blank" = class=3D"">http://lists.ovirt.org/mailman/listinfo/devel</a><br = class=3D""> </blockquote></div><br class=3D""></div></div> _______________________________________________<br class=3D"">Devel = mailing list<br class=3D""><a href=3D"mailto:Devel@ovirt.org" = class=3D"">Devel@ovirt.org</a><br = class=3D"">http://lists.ovirt.org/mailman/listinfo/devel</div></blockquote=
</div><br class=3D""></body></html>=
--Apple-Mail=_643EEE38-B765-4581-8070-B2E414E4F79B--
On 11/12/15 12:31 +0100, Michal Skrivanek wrote:
On 10 Dec 2015, at 16:36, Yaniv Kaul <ykaul@redhat.com> wrote:
On Thu, Dec 10, 2015 at 5:07 PM, Martin Polednik <mpolednik@redhat.com <mailto:mpolednik@redhat.com>> wrote: Hello developers,
tl;dr version: * deprecate report_host_threads_as_cores * remove cpuSockets, use sum(numaNodes.keys()) * report threadsPerCore for ppc64le / report total number of threads for ppc64le * work on our naming issues
I've been going over our capabilities reporting code in VDSM due to specific threading requirements on ppc64le platform and noticed few issues. Before trying to fix something that "works", I'm sending this mail to start a discussion regarding current and future state of the code.
First thing is the terminology. What we consider cpu sockets, cores and threads are in fact NUMA cells, sum of cores present in NUMA nodes and the same for threads. I'd like to see the code moving in a direction that is correct in this sense.
Note that I think users are more familiar with sockets-cores-threads than NUMA cells, terminology-wise.
we do report numa separately today, and we should keep doing that. I consider it another level of detail/complexity which many users do not care about. So we should keep both
The issue is not removing one or the other, but rather that what we report as CPU sockets/cores/threads are an actual NUMA sockets/cores/threads. As long as 1 CPU == 1 NUMA cell we're fine, but the POWER8 CPUs are 1 chip (socket) = 4 cores = 4 NUMA cells - reporting 4 as the number of sockets per cpu.
More important are the actual calculations. I believe we should draw an uncrossable line between cores and threads and not interfere with it at least on VDSM's side. That would mean deprecating report_host_threads_as_cores option. The actual algorithm used at present does calculate the numa cores and numa threads correctly given that there are no offline CPUs - most likely fine enough. We don't have to report the actual number of sockets though, as it is reported in numa* keys.
There is a reason for report_host_threads_as_cores option. I don't remember it right now, but it had to do with some limitation of some OS or license or something. I don't think we should deprecate it.
the idea was to remove that option from VDSM conf (as it’s cumbersome to use), and rather report all relevant information so engine can decide later on whether to count it this way or another Today it’s used as a simple “core multiplier” if your workload is running "good enough” in parallel on 2 threads within one core, we just consider it as additional available “cpu”. For some workloads where this assumption is not working well, and also for licensing or any other reason, you can disable it and see “half” of the cpus on x86 despite having HT enabled in BIOS.
On PPC this is more tricky as MArtin says below - (threads are not able to run multiple VMs simultaneously) - so we need to push that decision from vdsm up the chain.
It does fail to provide us with information that can be used in ppc64le environment, where for POWER8 we want to run the host without SMT while VMs would have multiple CPUs assigned. There are various configurations of so-called subcores in POWER8, where each CPU core can contain 1, 2 or 4 subcores. This configuration must be taken in consideration as given e.g. 160 threads overall, it is possible to run either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2 mode. We have to report either the total number of threads OR just the threadsPerCore setting, so the users know how many "CPUs" should be assigned to machines for optimal performance.
x per y sounds best to me but I think it’s even more complicated, if we consider offline CPUs(we don’t do that today) then the default picture on POWER8 currently looks like 20 cores in 4 numa cells, 8 threads per core. SMT is disabled altogether, so CPUs 1-7,9-15,… are offline. So should we report them or not? On x86 I would not do that as they are administratively disabled and can’t be used, however on ppc since RHEL 7.2 they are dynamically enabled on demand (if the guest topology uses threads as well), so they should be reported as available (or "sort-of-available”:)
If we report threads per core, the offline CPUs can be calculated from the available online CPUs and given threads per core value.
still, I think we should go with simple "sockets, cores/socket, threads/core” numbers, the rest needs to be computed or chosen from, based on additional detailed report of NUMA topology and online/offline CPU status perhaps with different behavior/capabilities on x86 and on power
YAY... do we have a comparison what libvirt knows / looks at (or they ignore it altogether?) Y.
As always, I welcome any opinions regarding the proposed ideas. Also note that all of the changes can be done via deprecation to be fully backwards compatible - except for the ppc part.
Regards, mpolednik _______________________________________________ Devel mailing list Devel@ovirt.org <mailto:Devel@ovirt.org> http://lists.ovirt.org/mailman/listinfo/devel <http://lists.ovirt.org/mailman/listinfo/devel>
_______________________________________________ Devel mailing list Devel@ovirt.org http://lists.ovirt.org/mailman/listinfo/devel
Top posting. Martin, I believe you have several important points. Since I was responsible for adding NUMA and the threading support in the engine (and yes, the vdsm conf is deprecated for several versions now), I can tell you that what we had up until now was sufficient to our needs. Since we have PPC and other technologies, we should give this a proper review. For example, with Cluster-on-die you can have 2 NUMA nodes on a single socket (we're about to order it to our lab) and AMD allows similar concepts in some of their processors these days. So what I suggest it that we'll review it together is order to provide a coherent representation which works for both architectures. I'd like Roy Golan to work on this, as it has direct implications on NUMA, CPU pinning, scheduling and other SLA related features. I'll arrange a meeting and once we have all the information we'll publish it in a wiki page. Thanks for bringing this up, Doron On Fri, Dec 11, 2015 at 3:20 PM, Martin Polednik <mpolednik@redhat.com> wrote:
On 11/12/15 12:31 +0100, Michal Skrivanek wrote:
On 10 Dec 2015, at 16:36, Yaniv Kaul <ykaul@redhat.com> wrote:
On Thu, Dec 10, 2015 at 5:07 PM, Martin Polednik <mpolednik@redhat.com <mailto:mpolednik@redhat.com>> wrote: Hello developers,
tl;dr version: * deprecate report_host_threads_as_cores * remove cpuSockets, use sum(numaNodes.keys()) * report threadsPerCore for ppc64le / report total number of threads for ppc64le * work on our naming issues
I've been going over our capabilities reporting code in VDSM due to specific threading requirements on ppc64le platform and noticed few issues. Before trying to fix something that "works", I'm sending this mail to start a discussion regarding current and future state of the code.
First thing is the terminology. What we consider cpu sockets, cores and threads are in fact NUMA cells, sum of cores present in NUMA nodes and the same for threads. I'd like to see the code moving in a direction that is correct in this sense.
Note that I think users are more familiar with sockets-cores-threads than NUMA cells, terminology-wise.
we do report numa separately today, and we should keep doing that. I consider it another level of detail/complexity which many users do not care about. So we should keep both
The issue is not removing one or the other, but rather that what we report as CPU sockets/cores/threads are an actual NUMA sockets/cores/threads. As long as 1 CPU == 1 NUMA cell we're fine, but the POWER8 CPUs are 1 chip (socket) = 4 cores = 4 NUMA cells - reporting 4 as the number of sockets per cpu.
More important are the actual calculations. I believe we should draw an uncrossable line between cores and threads and not interfere with it at least on VDSM's side. That would mean deprecating report_host_threads_as_cores option. The actual algorithm used at present does calculate the numa cores and numa threads correctly given that there are no offline CPUs - most likely fine enough. We don't have to report the actual number of sockets though, as it is reported in numa* keys.
There is a reason for report_host_threads_as_cores option. I don't remember it right now, but it had to do with some limitation of some OS or license or something. I don't think we should deprecate it.
the idea was to remove that option from VDSM conf (as it’s cumbersome to use), and rather report all relevant information so engine can decide later on whether to count it this way or another Today it’s used as a simple “core multiplier” if your workload is running "good enough” in parallel on 2 threads within one core, we just consider it as additional available “cpu”. For some workloads where this assumption is not working well, and also for licensing or any other reason, you can disable it and see “half” of the cpus on x86 despite having HT enabled in BIOS.
On PPC this is more tricky as MArtin says below - (threads are not able to run multiple VMs simultaneously) - so we need to push that decision from vdsm up the chain.
It does fail to provide us with information that can be used in ppc64le environment, where for POWER8 we want to run the host without SMT while VMs would have multiple CPUs assigned. There are various configurations of so-called subcores in POWER8, where each CPU core can contain 1, 2 or 4 subcores. This configuration must be taken in consideration as given e.g. 160 threads overall, it is possible to run either 20 VMs in smt8 mode, 40 VMs in smt4 mode or 80 VMs in smt2 mode. We have to report either the total number of threads OR just the threadsPerCore setting, so the users know how many "CPUs" should be assigned to machines for optimal performance.
x per y sounds best to me but I think it’s even more complicated, if we consider offline CPUs(we don’t do that today) then the default picture on POWER8 currently looks like 20 cores in 4 numa cells, 8 threads per core. SMT is disabled altogether, so CPUs 1-7,9-15,… are offline. So should we report them or not? On x86 I would not do that as they are administratively disabled and can’t be used, however on ppc since RHEL 7.2 they are dynamically enabled on demand (if the guest topology uses threads as well), so they should be reported as available (or "sort-of-available”:)
If we report threads per core, the offline CPUs can be calculated from the available online CPUs and given threads per core value.
still, I think we should go with simple "sockets, cores/socket,
threads/core” numbers, the rest needs to be computed or chosen from, based on additional detailed report of NUMA topology and online/offline CPU status perhaps with different behavior/capabilities on x86 and on power
YAY... do we have a comparison what libvirt knows / looks at (or they ignore it altogether?) Y.
As always, I welcome any opinions regarding the proposed ideas. Also note that all of the changes can be done via deprecation to be fully backwards compatible - except for the ppc part.
Regards, mpolednik _______________________________________________ Devel mailing list Devel@ovirt.org <mailto:Devel@ovirt.org> http://lists.ovirt.org/mailman/listinfo/devel < http://lists.ovirt.org/mailman/listinfo/devel>
_______________________________________________ Devel mailing list Devel@ovirt.org http://lists.ovirt.org/mailman/listinfo/devel
_______________________________________________ Devel mailing list Devel@ovirt.org http://lists.ovirt.org/mailman/listinfo/devel
participants (5)
-
Doron Fediuck -
Francesco Romani -
Martin Polednik -
Michal Skrivanek -
Yaniv Kaul