On Thu, Dec 19, 2013 at 12:36:30PM +0200, Abel Gordon wrote:
> On Thu, Dec 19, 2013 at 12:13 PM, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > On Thu, Dec 19, 2013 at 08:40:44AM +0200, Abel Gordon wrote:
> >> On Wed, Dec 18, 2013 at 12:43 PM, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >> > On Tue, Dec 17, 2013 at 12:04:42PM +0200, Razya Ladelsky wrote:
> >> >> Hi,
> >> >>
> >> >> Thank you all for your comments.
> >> >> I'm sorry for taking this long to reply, I was away on vacation..
> >> >>
> >> >> It was a good, long discussion, many issues were raised, which we'd like
> >> >> to address with the following proposed roadmap for Elvis patches.
> >> >> In general, we believe it would be best to start with patches that are
> >> >> as simple as possible, providing the basic Elvis functionality,
> >> >> and attend to the more complicated issues in subsequent patches.
> >> >>
> >> >> Here's the road map for Elvis patches:
> >> >
> >> > Thanks for the follow up. Some suggestions below.
> >> > Please note they suggestions below merely represent
> >> > thoughts on merging upstream.
> >> > If as the first step you are content with keeping this
> >> > work as out of tree patches, in order to have
> >> > the freedom to experiment with interfaces and
> >> > performance, please feel free to ignore them.
> >> >
> >> >> 1. Shared vhost thread for multiple devices.
> >> >>
> >> >> The way to go here, we believe, is to start with a patch having a shared
> >> >> vhost thread for multiple devices of the SAME vm.
> >> >> The next step/patch may be handling vms belonging to the same cgroup.
> >> >>
> >> >> Finally, we need to extend the functionality so that the shared vhost
> >> >> thread
> >> >> serves multiple vms (not necessarily belonging to the same cgroup).
> >> >>
> >> >> There was a lot of discussion about the way to address the enforcement
> >> >> of cgroup policies, and we will consider the various solutions with a
> >> >> future
> >> >> patch.
> >> >
> >> > With respect to the upstream kernel,
> >> > I'm not sure a bunch of changes just for the sake of guests with
> >> > multiple virtual NIC cards makes sense.
> >> > And I wonder how this step, in isolation, will affect e.g.
> >> > multiqueue workloads.
> >> > But I guess if the numbers are convincing, this can be mergeable.
> >>
> >> Even if you have a single multiqueue device this change allows
> >> to create one vhost thread for all the queues, one vhost thread per
> >> queue or any other combination. I guess that depending on the workload
> >> and depending on the system utilization (free cycles/cores, density)
> >> you would prefer
> >> to use one or more vhost threads.
> >
> > That is already controllable from the guest though, which likely has a better
> > idea about the workload.
>
> but the guest has no idea about what's going on in the host system
> (e.g. other VMs I/O, cpu utilization of the host cores...)
But again, you want to do things per VM now so you will have no idea
about other VMs, right? Host cpu utilization could be a useful input
for some heuristics, I agree, but nothing prevents us from sending
this info to guest agent and controlling multiqueue based on that
(kind of like balloon).
> >
> >> >
> >> >>
> >> >> 2. Creation of vhost threads
> >> >>
> >> >> We suggested two ways of controlling the creation and removal of vhost
> >> >> threads:
> >> >> - statically determining the maximum number of virtio devices per worker
> >> >> via a kernel module parameter
> >> >> - dynamically: Sysfs mechanism to add and remove vhost threads
> >> >>
> >> >> It seems that it would be simplest to take the static approach as
> >> >> a first stage. At a second stage (next patch), we'll advance to
> >> >> dynamically
> >> >> changing the number of vhost threads, using the static module parameter
> >> >> only as a default value.
> >> >
> >> > I'm not sure how independent this is from 1.
> >> > With respect to the upstream kernel,
> >> > Introducing interfaces (which we'll have to maintain
> >> > forever) just for the sake of guests with
> >> > multiple virtual NIC cards does not look like a good tradeoff.
> >> >
> >> > So I'm unlikely to merge this upstream without making it useful cross-VM,
> >> > and yes this means isolation and accounting with cgroups need to
> >> > work properly.
> >>
> >> Agree, but even if you use a single multiqueue device having the
> >> ability to use 1 thread to serve all the queues or multiple threads to
> >> serve all the queues looks like a useful feature.
> >
> > Could be. At the moment, multiqueue is off by default because it causes
> > regressions for some workloads as compared to a single queue.
> > If we have heuristics in vhost that fix this by auto-tuning threading, that
> > would be nice. But if you need to tune it manually anyway,
> > then from upstream perspective it does not seem to be worth it - you can just
> > turn multiqueue on/off in the guest.
>
> I see. But we are mixing again between the policy and the mechanism.
> We first need a mechanism to control the system and then we need to
> implement the policy to orchestrate it (whenever it will be
> implemented in the kernel as part of vhost or outside in user-space).
> I don't see why to wait to have a policy to upstream the mechanism. If
> we upstream the mechanism in a manner that the defaults do not affect
> today's vhost behavior, then it will be possible to play with the
> policies and upstream them later.
Well it all hinges on whether it's in userspace actually.
Interfaces to userspace must be maintained forever, there's no way to know
they are right and make sense before they are used, so we must not merge
the interface before there is userspace using it correctly.
If it's in-kernel, we can merge a bit of unused code into kernel
just to make it easier for you to make progress, provided it's
well isolated and doesn't make life much harder for others
working on same code-base.
> >
> >
> >> >
> >> >> Regarding cwmq, it is an interesting mechanism, which we need to explore
> >> >> further.
> >> >> At the moment we prefer not to change the vhost model to use cwmq, as some
> >> >> of the issues that were discussed, such as cgroups, are not supported by
> >> >> cwmq, and this is adding more complexity.
> >> >> However, we'll look further into it, and consider it at a later stage.
> >> >
> >> > Hmm that's still assuming some smart management tool configuring
> >> > this correctly. Can't this be determined automatically depending
> >> > on the workload?
> >> > This is what the cwmq suggestion was really about: detect
> >> > that we need more threads and spawn them.
> >> > It's less about sharing the implementation with workqueues -
> >> > would be very nice but not a must.
> >>
> >> But how cwmq can consider cgroup accounting ?
> >
> > I think cwmq is just a replacement for our own thread pool.
> > It doesn't make cgroup accounting easier or harder.
>
> cwmq doesn't solve the cgroup issue, that's the problem. IMHO, it will
> be much simpler consider cgrouups with our own threading model.
Bandan here arrived at the same conclusion.
> >
> >> In any case, IMHO, the kernel should first provide the "mechanism" so
> >> later on a user-space management application (the "policy") can
> >> orchestrate it.
> >
> > I think policy would be something coarse-grained, like setting priority.
> > Making detailed scheduling decisions in userspace seems wrong somehow:
> > what does management application know that kernel doesn't?
>
> The code of the management application can be
> changed/modified/customized... also the management application can
> consider additional inputs (like per VM SLAs) that are not available
> to the kernel.
An alternative is to express these in terms of cgroups.
> Think about MOM (memory over-commit manager) but for vhost (vhost I/O
> manager ?VHIOM ?)
I worry that CPU and network load is much more dynamic than
memory load, and that in real life userspace simply won't be able to
react to changes fast enough.
Also that this will interact with scheduler in strange ways, e.g.
manager wants to free up the CPU and moves some vhost threads
off it, scheduler sees there's free CPU and moves some more jobs there.
After all it might be the best we can do for over-committed systems,
but it seems that it will have to be inter-VM to be really useful.
Overall one of the main points of vhost is a single interface
that let us access all kind of kernel APIs in a uniform way:
we tell kernel what we want it to do, and let it get about
executing that in the most efficient manner possible.
If you insist on managing everything from userspace you lose
some of these advantages.
That's why I keep suggesting we try to give scheduler more hints
about our expected behaviour.
> >
> >
> >> >
> >> >
> >> >
> >> >> 3. Adding polling mode to vhost
> >> >>
> >> >> It is a good idea making polling adaptive based on various factors such as
> >> >> the I/O rate, the guest kick overhead(which is the tradeoff of polling),
> >> >> or the amount of wasted cycles (cycles we kept polling but no new work was
> >> >> added).
> >> >> However, as a beginning polling patch, we would prefer having a naive
> >> >> polling approach, which could be tuned with later patches.
> >> >>
> >> >
> >> > While any polling approach would still need a lot of testing to prove we
> >> > don't for example steal CPU from guest which could be doing other useful
> >> > work, given that an exit is at least 1.5K cycles at least in theory it
> >> > seems like something that can improve performance. I'm not sure how
> >> > naive we can be without introducing regressions for some workloads.
> >> > For example, if we are on the same host CPU, there's no
> >> > chance busy waiting will help us make progress.
> >> > How about detecting that the VCPU thread that kicked us
> >> > is currently running on another CPU, and only polling in
> >> > this case?
> >> >
> >> >> 4. vhost statistics
> >> >>
> >> >> The issue that was raised for the vhost statistics was using ftrace
> >> >> instead of the debugfs mechanism.
> >> >> However, looking further into the kvm stat mechanism, we learned that
> >> >> ftrace didn't replace the plain debugfs mechanism, but was used in
> >> >> addition to it.
> >> >>
> >> >> We propose to continue using debugfs for statistics, in a manner similar
> >> >> to kvm,
> >> >> and at some point in the future ftrace can be added to vhost as well.
> >> >
> >> > IMHO which kvm stat is a useful script, the best tool
> >> > for perf stats is still perf. So I would try to integrate with that.
> >> > How it works internally is IMHO less important.
> >> >
> >> >> Does this plan look o.k.?
> >> >> If there are no further comments, I'll start preparing the patches
> >> >> according to what we've agreed on thus far.
> >> >> Thank you,
> >> >> Razya
> >> >
> >> > I think a good place to try to start merging upstream would be 3 and 4.
> >> > So if you want to make it easier to merge things upstream, try to keep 3
> >> > and 4 independent from 1 and 2.
> >>
> >> Note -1- and -3- are strongly related. If you have a thread that
> >> serves multiple queues (whenever they belong to the same device/vm or
> >> not) then this thread will be polling multiple queues at the same
> >> time. This increases the chances you will find pending work to do in
> >> some queue. In other words, you reduce the cycles wasted for polling.
> >> In the other hand, if you run multiple threads and these threads do
> >> polling simultaneously then the threads may starve each other and
> >> reduce performance (if they are scheduled to run in the same core). In
> >> addition, a shared thread can decide when it should stop processing a
> >> given queue and switch to other queue because by polling the thread
> >> knows when new requests were added to a queue (this is what we called
> >> fine-grained I/O scheduled heuristics)
> >>
> >> So, seems like polling makes more sense when you serve multiple queues
> >> with the same thread.
> >>
> >> Abel.
> >
> > A combination might bring gains in more workloads, but it should work on its
> > own too. It's quite possible that only a single VM is active, others are
> > idle. So either polling should handle that well or be smart enough to turn
> > itself off in this case.
>
> Agree, but we are discussing again about the policy. We first need
> the mechanism :)
>
> Abel.
So you want to control this from a management application as well?
That's a very low level detail to expose to management.
Why do we ever want to poll? It burns CPU uselessly.
There are several reasons for polling to be good:
1. scheduler overhead is too large, we are thrashing
between multiple threads running on the same CPU.
This is the one you are thinking of I guess: it can
be controlled from the management application.
So here some management interface might make some sense.
But to know how long to poll you would need to know
what the scheduler overhead is, then poll for
a fraction of that.
2. KVM exit overhead is too large, we are wasting cycles
on exit/reentry.
Userspace doesn't easily know about exits though.
Again, we'd need to measure how long does an exit take,
then poll for a fraction of that.
This one is easier to measure from guest, so maybe
we should let guest control this.
3. If we halt the CPU it might take too long to wake up.
If we expect to be woken up soon we'd be better off
busy waiting.
This looks like it would be better addressed by an in-kernel API
that says "I expect to be woken up in X microseconds".
This way if some other event occurs instead, it will be
handled instead of waiting for us to finish polling.
As you see there are many factors in play here.
That's why we can't just add userspace interface "poll for X cycles":
to support item 1. maintaining this interface would block
support for items 2 and 3.
> >
> >> >
> >> > Thanks again,
> >> >
> >> > --
> >> > MST
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
