2:53 a.m.
On Tue, Jun 12, 2018 at 10:11:30AM +0000, Wang, Huaqiang wrote:
Hi Martin,
Thanks for your comments, please see my update inline below.
> -----Original Message-----
> From: Martin Kletzander [mailto:mkletzan@redhat.com]
> Sent: Monday, June 11, 2018 4:30 PM
> To: Wang, Huaqiang <huaqiang.wang(a)intel.com>
> Cc: libvir-list(a)redhat.com; Feng, Shaohe <shaohe.feng(a)intel.com>; Niu, Bing
> <bing.niu(a)intel.com>; Ding, Jian-feng <jian-feng.ding(a)intel.com>; Zang,
Rui
> <rui.zang(a)intel.com>
> Subject: Re: [libvirt] [RFC PATCH 0/3] RFC for X86 RDT Cache Monitoring
> Technology (CMT) support
>
> [It would be nice if you wrapped the long lines]
I'll pay attention to these long lines. Thanks for advices.
No need to, most email clients can do that automatically. Doing stuff like this
manually is very unproductive :).
>
> On Fri, Jun 08, 2018 at 05:02:16PM +0800, Wang Huaqiang wrote:
> >This is an RFC request for supporting CPU Cache Monitoring Technology (CMT)
> feature in libvirt. Since MBM is also another feature which is very close to CMT,
> for simplicity we only discuss CMT here. MBM is the followup that will be
> implemented after CMT.
> >About CMT please refer to Intel x86 SDM section 17.18 of volume 3
> (link:https://software.intel.com/en-us/articles/intel-sdm).
> >
>
> Can you elaborate on how is this different to the CMT perf event that is already
> in libvirt and can be monitored through domstats API?
Due to kernel interface removal of the perf events 'cmt,mbmt,mbml', the libvirt
will no
longer work with latest kernel. Please examine following link for details.
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/c...,
This serials is trying to provide the similar functions of this missing part for
reporting
cmt, mbmt and mbml information. First we only focus on cmt.
Comparing with 'CMT perf event already in libvirt', I am trying to implement
almost
the same output as 'perf.cmt' in the output message of 'domstats', but
with another
name , such as 'resctrl.cmt' or 'rdt.cmt' (or some others).
Another difference is that the underlying implementation is done through the
kernel resctrl fs.
This serials also attempts to provide a command interface for enabling and disabling
cmt feature in scope of whole domain as original perf event based cmt could be
controlled, enabled or disabled, through specifying '--enable cmt' or
'--disable cmt'
while invoking command 'virsh perf <domain>'.
Our version is like 'virsh resctrl <domain> --enable' with a difference of
no suffix of
'cmt'. The 'cmt' is omitted because the CMT and MBM function are both
enabled
whenever a valid resctrl fs sub-folder created, there is no way to disable one while
enable another one, such as enabling CMT while disabling MBML at the same time.
This serials is trying to stick to interfaces exposed by perf event based CMT/MBM
and provide an interface substitution for perf event based CMB/MBM, such as
the perf based CMT only provides the cache occupancy information for whole
domain only. We are also in thinking providing the capability to provide the
cache occupancy information based on vcpus groups which may be specified in
XML file.
For example, if we have following configuration:
<cputune>
<vcpupin vcpu='0' cpuset='1'/>
<vcpupin vcpu='1' cpuset='3-4'/>
<vcpupin vcpu='2' cpuset='4-5'/>
<vcpupin vcpu='3' cpuset='6-7'/>
<cachetune vcpus='0'>
<cache id='0' level='3' type='both' size='2816'
unit='KiB'/>
<cache id='1' level='3' type='both' size='2816'
unit='KiB'/>
</cachetune>
<cachetune vcpus='1-2'>
<cache id='0' level='3' type='both' size='2816'
unit='KiB'/>
<cache id='1' level='3' type='both' size='2816'
unit='KiB'/>
</cachetune>
<rdt-monitoring vcpu='0' enable='yes'>
<rdt-monitoring vcpu='1-2' enable='yes'>
<rdt-monitoring vcpu='3' enable='yes'>
</cputune>
The 'domstats' will output following information regarding cmt
[root@dl-c200 libvirt]# virsh domstats vm1 --resctrl
Domain: 'vm1'
rdt.cmt.total=645562
rdt.cmt.vcpu0=104331
rdt.cmt.vcpu1_2=203200
rdt.cmt.vcpu3=340129
beware as 1-4 is something else than 1,4 so you need to differentiate that. Or
to make it easier to parse for consumers of that API just list each vcpu on its
own line (but then you need to say which are counted together). Or group them:
rdt.cmt.total=645562
rdt.cmt.group0.value=104331
rdt.cmt.group0.vcpus=0
rdt.cmt.group0.value=203200
rdt.cmt.group0.vcpus=1-2
rdt.cmt.group0.value=340129
rdt.cmt.group0.vcpus=3
Honestly, I don't care that much how it is going to look, but it needs to be
easy to parse and understand.
Those updates address your comment for " Can you elaborate on how
is
this different to the CMT perf event that is already in libvirt and can be
monitored through domstats API?", any input is welcome.
Great, this information (or rather a brief summary) should be part of the patch
series. Not necessarily the commit messages (some of the things would fit
there), but at least the cover letter. Otherwise you might get the same
question next time and will have to provide the same answer to the next reviewer
and so on.
>
> https://libvirt.org/formatdomain.html#elementsPerf
>
> >## About '_virResctrlMon' interface
> >
> >The cache allocation technology (CAT) has already been implemented in
> >util/virresctrl.* which interacts with Linux kernel resctrl file
> >system. Very simlimar to CAT, the CMT object is represented by 'struct
> >_virResctrlMon', which is
> >
> >```
> >struct _virResctrlMon {
> > virObject parent;
> >
> > /* pairedalloc: pointer to a resctrl allocaion it paried with.
> > * NULL for a resctrl monitoring group not associated with
> > * any allocation. */
> > virResctrlAllocPtr pairedalloc;
> > /* The identifier (any unique string for now) */
> > char *id;
> > /* libvirt-generated path, may be identical to alloction path
> > * may not if allocation is ready */
> > char *path;
> >};
> >```
> >
> >Almost following the same logic behind '_virResctrlAlloc' which is
mainly
> presented in file 'virresctrl.c', a group of APIs has been designed to
manipulate
> '_virResctrlMon'. The '_virResctrlMon' shares a lot in common with
> '_virResctrlAlloc' except field 'pairedalloc'.
> >'pairedalloc' stores the pointer of paired resctrl allocation object.
With current
> libvirt resctrl implementation, if a resctrl '_virResctrlAlloc' object is
created, the
> CMT hardware is enabled automatically and shares the same folder under same
> resctrlfs, I call the CMT '_virResctrlMon' object that shares the same
folder
> under resctrlfs as 'paired' _virResctrlMon, further, one
'_virResctrlMon' and one
> '_virResctrlAlloc' are a pare. In '_virResctrlMon' the paired
'_virResctrlAlloc' is
> tracked through pairedalloc. paired mon group could not be dynamically enabled
> or disabled during runtime.
> >'pairedalloc' could be set to NULL, which creates a non-paired mon group
> object. Which is necessory because CMT could work independently to monitor
> the utilization of critical CPU resouces (cache or memory bandwidth) without
> allocating any dedicated cache or memory bandwidth. A non-paired mon group
> object represents an independent working CMT. Non-paired mon group could be
> enabled or disabled during runtime.
> >
> >## About virsh command 'resctrl'
> >
> >To set or get the resctrl mon group (hardware CMT), a virsh command
'resctrl'
> is created. here are the common usages:
>
> The command does make sense for people who know how the stuff works on
> the inside or have seen the code in libvirt. For other users the name
'resctrl' is
> going to feel very much arbitrary. We re trying to abstract the details for users,
> so I don't see why it should be named 'resctrl' when it handles "RDT
Monitoring
> Status".
Agree. 'resctrl' do make a lot of confusion to end users.
Since the underlying kernel interface combines CAT and MBM features together,
what I mean is , the files ' llc_occupancy', ' mbm_local_bytes' and '
mbm_total_bytes'
that represent the information of cache, local memory bandwidth, and total
memory bandwidth respectively are created automatically and simultaneously for
each resctrl group, there is no way to enable one and disable another one. So for
a command which affects both cache and memory bandwidth, I would like to use
the word 'rdt' as the key command word. Both cache monitoring(CMT) and memory
bandwidth monitoring(MBM) are belong to the scope of RDT monitoring.
So to replace the confusing word 'resctrl', I'd like to use 'rdtmon'
as command name,
the command 'virsh resctrl <domain>' would be changed to 'virsh rdtmon
<domain>'.
Also, here welcoming any suggestions from community.
Libvirt tries to abstract various vendor-specific things. For example AMD's SEV
is abstracted under the name `launch-security` IIRC so that if there are more in
the future not all the code needs to be duplicated. In the same sense Intel's
RDT could be named in a more generic sense. Resource Control and Monitoring
seems to reflect what it does, but it's kind of a mouthful. Maybe others will
have better ideas. I'm bad at naming.
>
> >```
> >[root@dl-c200 david]# virsh list --all
> > Id Name State
> >----------------------------------------------------
> > 1 vm3 running
> > 3 vm2 running
> > - vm1 shut off
> >```
> >
> >### Test on a running domain vm3
> >To get RDT monitoring status, type 'virsh resctrl <domain>'
> >```
> > [root@dl-c200 david]# virsh resctrl vm3
> > RDT Monitoring Status: Enabled
> >```
> >
> >To enable RDT monitoring, type 'virsh resctrl <domain> --enable'
> >```
> > [root@dl-c200 david]# virsh resctrl vm3 --enable
> > RDT Monitoring Status: Enabled
> >```
> >
> >To diable RDT monitoring, type 'virsh resctrl <domain> --disable'
> >```
> > [root@dl-c200 david]# virsh resctrl vm3 --disable
> > RDT Monitoring Status: Disabled
> >
> > [root@dl-c200 david]# virsh resctrl vm3
> > RDT Monitoring Status: Disabled
> >```
> >
> >### test on domain not running vm1
> >if domain is not active, it will fail to set RDT monitoring status, and also get
the
> state of 'disabled'
> >```
> > [root@dl-c200 david]# virsh resctrl vm1
> > RDT Monitoring Status: Disabled
> >
> > [root@dl-c200 david]# virsh resctrl vm1 --enable
> > error: Requested operation is not valid: domain is not running
> >
> > [root@dl-c200 david]# virsh resctrl vm1 --disable
> > error: Requested operation is not valid: domain is not running ```
> >
>
> Can't these commands enable it in the XML? It would be nice if the XML part
> was
> shown here in the explanation.
In the POC code of the first version there is no XML changes, and could not be
enabled/disabled through XML file.
Let's have a discuss and add this function, how about this configuration
<cputune>
<cachetune vcpus='1-2'>
<cache id='0' level='3' type='both' size='2816'
unit='KiB'/>
<cache id='1' level='3' type='both' size='2816'
unit='KiB'/>
</cachetune>
<rdt-monitoring vcpu='0' enable='no'>
<rdt-monitoring vcpu='1-2' enable='yes'>
<rdt-monitoring vcpu='3' enable='yes'>
</cputune>
Just so we are on the same note, it doesn't have to have an option to be
enabled/disabled in the XML. However, you probably still need to keep the state
of that information somewhere across libvirtd restarts. If there is any, I
haven't gone through the code.
With upper setting,
- Two rdt monitoring groups will be created along with the launch of vm.
- <rdt-monitoring vcpu='1-2' enable='yes'> is created automatically
due to the
setting of <cachetune>. Under resctrl fs, the resctrl allocation and rdt monitoring
group are presented in the way of sub-folders, we cannot create two sub-folders
under resctrl fs folders for one process. so a resctrl allocation will create a rdt
monitoring group as well. This rdt monitoring group could not be disabled in
runtime because there is no way to disable resctrl allocation (CAT) in runtime.
- <rdt-monitoring vcpu='3' enable='yes'> creates another default
enabled rdt
monitoring group, and task id (pid associated with vcpu3) will be put into the
'tasks' file. This rdt monitoring over vcpu 3 could be enabled or disabled in
runtime through command such as 'virsh rdtmon --enable vcpu3' .
The MBM feature will also be enabled or disabled with this command.
- <rdt-monitoring vcpu='0' enable='no'> specifies the default CAT
state for vcpu0
of domain, which is disabled after launch, and could be changed in runtime.
There are many places where stuff can be created. I started going down the
rabbit hole again (like last time when I was implementing CAT) and again, the
kernel interface is horrible. Inconsistent naming, poor documentation (or maybe
I'm just a very bad reader). I hope someone will join this review because I
can't sensibly map the kernel interface to whatever libvirt might do/expose. I
already wasted so much time on CAT and I don't want to go back to that again.
Let's not do any XML changes unless we find out they are actually needed.
> >### test on domain vm2
> >domain vm2 is active and the CAT functionality is enabled through
'cachetune'
> (configured in 'cputune/cachetune' section). So the resctrl mon group is a
> 'paried' one, for 'pared' mon group, the RDT monitoring could not be
disabled. If
> it is allowed to disable 'paire' mon group, we have to destroy resctrl
allocation
> folders which is not supported by current cache allocation design.
>
> What if you have multiple cachetunes? What if the cachetune is only set for one
> vcpu and you want to monitor the others as well? I guess I have to see the
> patches to understand why you have so much information stored for something
> that
> looks like a boolean (enable/disable).
At the time I raised this RFC, there is no design for reporting rdt monitoring
information in granularity of cachetune, only report cache /memory bandwidth
information for whole domain.
But now I'd like to discuss the design that I list above, reporting rdt monitoring
Information based on the setting of rdt-monitoring(cachetune) groups. Need your
comments.
I just wanted to know what is the preferred approach. If we're creating
mon_groups/domain_name_vcpus_X/ or just new resctrl group (there is not much of
a difference in that). Does it take hot-(un)plug of vcpus into consideration?
How about emulator threads and iothreads? I know libvirt doesn't support them
yet for CAT, but that'd be a good way to start adding features to libvirt IMHO.
Or live changes to cachetunes. If we have that, then maybe the addition of
monitoring will make more sense and it will fit more nicely (since we'll have a
more complete picture).
>
> >```
> > [root@dl-c200 libvirt]# virsh resctrl vm2 --enable
> > RDT Monitoring Status: Enabled (forced by cachetune)
> >
> > [root@dl-c200 libvirt]# virsh resctrl vm2 --disable
> > RDT Monitoring Status: Enabled (forced by cachetune)
> >
> > [root@dl-c200 libvirt]# virsh resctrl vm2
> > RDT Monitoring Status: Enabled (forced by cachetune)
> >```
> >
> >## About showing the utilization information of RDT
> >
> >A domstats field has been created to show the utilization of RDT resources, the
> command is like this:
> >```
> > [root@dl-c200 libvirt]# virsh domstats --resctrl
> > Domain: 'vm1'
> > resctrl.cmt=0
> >
> > Domain: 'vm3'
> > resctrl.cmt=180224
> >
> > Domain: 'vm2'
> > resctrl.cmt=2613248
> >```
> >
> >
> >Wang Huaqiang (3):
> > util: add Intel x86 RDT/CMT support
> > tools: virsh: add command for controling/monitoring resctrl
> > tools: virsh domstats: show RDT CMT resource utilization information
> >
> > include/libvirt/libvirt-domain.h | 10 ++
> > src/conf/domain_conf.c | 28 ++++
> > src/conf/domain_conf.h | 3 +
> > src/driver-hypervisor.h | 8 +
> > src/libvirt-domain.c | 92 +++++++++++
> > src/libvirt_private.syms | 9 +
> > src/libvirt_public.syms | 6 +
> > src/qemu/qemu_driver.c | 189 +++++++++++++++++++++
> > src/qemu/qemu_process.c | 65 +++++++-
> > src/remote/remote_daemon_dispatch.c | 45 +++++
> > src/remote/remote_driver.c | 2 +
> > src/remote/remote_protocol.x | 28 +++-
> > src/remote_protocol-structs | 12 ++
> > src/util/virresctrl.c | 316 +++++++++++++++++++++++++++++++++++-
> > src/util/virresctrl.h | 44 +++++
> > tools/virsh-domain-monitor.c | 7 +
> > tools/virsh-domain.c | 74 +++++++++
> > 17 files changed, 933 insertions(+), 5 deletions(-)
> >
> >--
> >2.7.4
> >
> >--
> >libvir-list mailing list
> >libvir-list(a)redhat.com
> >https://www.redhat.com/mailman/listinfo/libvir-list