[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt][RFC PATCH] add a new 'default' option for attribute mode in numatune

On Tue, Aug 11, 2020 at 04:39:42PM +0800, Zhong, Luyao wrote:

On 8/7/2020 4:24 PM, Martin Kletzander wrote:
On Fri, Aug 07, 2020 at 01:27:59PM +0800, Zhong, Luyao wrote:

On 8/3/2020 7:00 PM, Martin Kletzander wrote:
On Mon, Aug 03, 2020 at 05:31:56PM +0800, Luyao Zhong wrote:
Hi Libvirt experts,

I would like enhence the numatune snippet configuration. Given a
example snippet:

   <memory mode="strict" nodeset="1-4,^3"/>
   <memnode cellid="0" mode="strict" nodeset="1"/>
   <memnode cellid="2" mode="preferred" nodeset="2"/>

Currently, attribute mode is either 'interleave', 'strict', or
I propose to add a new 'default'  option. I give the reason as

Presume we are using cgroups v1, Libvirt sets cpuset.mems for all vcpu
according to 'nodeset' in memory element. And translate the memnode
element to
qemu config options (--object memory-backend-ram) for per numa cell,
invoking mbind() system call at the end.[1]

But what if we want using default memory policy and request each guest
numa cell
pinned to different host memory nodes? We can't use mbind via qemu
config options,
because (I quoto here) "For MPOL_DEFAULT, the nodemask and maxnode
arguments must
be specify the empty set of nodes." [2]

So my solution is introducing a new 'default' option for attribute
mode. e.g.

   <memory mode="default" nodeset="1-2"/>
   <memnode cellid="0" mode="default" nodeset="1"/>
   <memnode cellid="1" mode="default" nodeset="2"/>

If the mode is 'default', libvirt should avoid generating qemu command
'--object memory-backend-ram', and invokes cgroups to set cpuset.mems
for per guest numa
combining with numa topology config. Presume the numa topology is :

   <cell id='0' cpus='0-3' memory='512000' unit='KiB' />
   <cell id='1' cpus='4-7' memory='512000' unit='KiB' />

Then libvirt should set cpuset.mems to '1' for vcpus 0-3, and '2' for
vcpus 4-7.

Is this reasonable and feasible? Welcome any comments.

There are couple of problems here.  The memory is not (always)
by the
vCPU threads.  I also remember it to not be allocated by the process,
but in KVM
in a way that was not affected by the cgroup settings.

Thanks for your reply. Maybe I don't get what you mean, could you give
me more context? But what I proposed will have no effect on other memory

Check how cgroups work.  We can set the memory nodes that a process will
allocate from.  However to set the node for the process (thread) QEMU
needs to
be started with the vCPU threads already spawned (albeit stopped).  And
for that
QEMU already allocates some memory.  Moreover if extra memory was allocated
after we set the cpuset.mems it is not guaranteed that it will be
allocated by
the vCPU in that NUMA cell, it might be done in the emulator instead or
the KVM
module in the kernel in which case it might not be accounted for the
actually causing the allocation (as we've already seen with Linux).  In all
these cases cgroups will not do what you want them to do.  The last case
be fixed, the first ones are by default not going to work.

That might be
fixed now,

But basically what we have against is all the reasons why we started
QEMU's command line arguments for all that.

I'm not proposing use QEMU's command line arguments, on contrary I want
using cgroups setting to support a new config/requirement. I give a
solution about if we require default memory policy and memory numa

And I'm suggesting you look at the commit log to see why we *had* to add
command line arguments, even though I think I managed to describe most
of them
above already (except for one that _might_ already be fixed in the
kernel).  I
understand the git log is huge and the code around NUMA memory
allocation was
changing a lot, so I hope my explanation will be enough.

Thank you for detailed explanation, I think I get it now. We can't
guarantee memory allocation matching requirement since there is a time
slot before setting cpuset.mems.

That's one of the things, although this one could be avoided (by setting a
global cgroup before exec()).

Sorry, but I think it will more likely break rather than fix stuff.
Maybe this
could be dealt with by a switch in `qemu.conf` with a huge warning above

I'm not trying to fix something, I propose how to support a new
requirement just like I stated above.

I guess we should take a couple of steps back, I don't get what you are
to achieve.  Maybe if you describe your use case it will be easier to
reach a

Yeah, I do have a usecase I didn't mention before. It's a feature in
kernel but not merged yet, we call it memory tiering.

If memory tiering is enabled on host, DRAM is top tier memory, and
PMEM(persistent memory) is second tier memory, PMEM is shown as numa
node without cpu. For short, pages can be migrated between DRAM and PMEM
based on DRAM pressure and how cold/hot they are.

We could configure multiple memory migrating path. For example,
node 0: DRAM, node 1: DRAM, node 2: PMEM, node 3: PMEM
we can make 0+2 to a group, and 1+3 to a group. In each group, page is
allowed to migrated down(demotion) and up(promotion).

If **we want our VMs utilizing memory tiering and with NUMA topology**,
we need handle the guest memory mapping to host memory, that means we
need bind each guest numa node to a memory nodes group(DRAM node + PMEM
node) on host. For example, guest node 0 -> host node 0+2.

However, only cgroups setting can make the memory tiering work, if we
use mbind() system call, demoted pages will never go back to DRAM.
That's why I propose to add 'default' option and bypass mbind in QEMU.

I hope I make myself understandable. I'll appreciate if you could give
some suggestion.

This comes around every couple of months/years and bites us in the back no
matter what way we go (every time there is someone who wants it the other way).
That's why I think there could be a way for the user to specify whether they
will likely move the memory or not and based on that we would specify
`host-nodes` and `policy` to qemu or not.  I think I even suggested this before
(or probably delegated it to someone else for a suggestion so that there is more
discussion), but nobody really replied.

So what we need, I think, is a way for someone to set a per-domain information
whether we should bind the memory to nodes in a changeable fashion or not.  I'd
like to have it in as well.  The way we need to do that is, probably,
per-domain, because adding yet another switch for each place in the XML where we
can select a NUMA memory binding would be a suicide.  There should also be no
need for this to be enabled per memory-(module, node), so it should work fine.

Ideally we'd discuss it with others, but I think I am only one of a few people
who dealt with issues in this regard.  Maybe Michal (Cc'd) also dealt with some
things related to the binding, so maybe he can chime in.


Have a nice day,





Attachment: signature.asc
Description: PGP signature

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]