Re: [PATCH] KVM: arm64: Limit stage2_apply_range() batch size to smallest block

[Krister Johansen]

Hi Marc,

On Sat, Mar 30, 2024 at 10:17:43AM +0000, Marc Zyngier wrote:
> On Fri, 29 Mar 2024 19:15:37 +0000,
> Krister Johansen <kjlx@xxxxxxxxxxxxxxxxxx> wrote:
> > On Fri, Mar 29, 2024 at 06:48:38AM -0700, Oliver Upton wrote:
> > > On Thu, Mar 28, 2024 at 12:05:08PM -0700, Krister Johansen wrote:
> > > > stage2_apply_range() for unmap operations can interfere with the
> > > > performance of IO if the device's interrupts share the CPU where the
> > > > unmap operation is occurring.  commit 5994bc9e05c2 ("KVM: arm64: Limit
> > > > stage2_apply_range() batch size to largest block") improved this.  Prior
> > > > to that commit, workloads that were unfortunate enough to have their IO
> > > > interrupts pinned to the same CPU as the unmap operation would observe a
> > > > complete stall.  With the switch to using the largest block size, it is
> > > > possible for IO to make progress, albeit at a reduced speed.
> > > 
> > > Can you describe the workload a bit more? I'm having a hard time
> > > understanding how you're unmapping that much memory on the fly in
> > > your workload. Is guest memory getting swapped? Are VMs being torn
> > > down?
> > 
> > Sorry I wasn't clear here.  Yes, it's the VMs getting torn down that's
> > causing the problems.  The container VMs don't have long lifetimes, but
> > some may be up to 256Gb in size, depending on the user.  The workloads
> > running the VMs aren't especially performance sensitive, but their users
> > do notice when network connections time-out.  IOW, if the performance is
> > bad enough to temporarily prevent new TCP connections from being
> > established or requests / responses being recieved in a timely fashion,
> > we'll hear about it.  Users deploy their services a lot, so there's a
> > lot of container vm churn.  (Really it's automation redeploying the
> > services on behalf of the users in response to new commits to their
> > repos...)
> 
> I think this advocates for a teardown-specific code path rather than
> just relying on the usual S2 unmapping which is really designed for
> eviction. There are two things to consider here:
> 
> - TLB invalidation: this should only take a single VMALLS12E1, rather
>   than iterating over the PTs
> 
> - Cache maintenance: this could be elided with FWB, or *optionally*
>   elided if userspace buys in a "I don't need to see the memory of the
>   guest after teardown" type of behaviour

This approach would work for this workload, I think.  The hardware
supports FWB and AFAIK isn't looking at the guest memory after teardown.
This is also desirable because in the future we'd like to support
hotplug of VFIO devices. A separate path for unmap the memory used by
the device vs unmap all of the guest seems smart.

> > > Also, it seems a bit odd to steer interrupts *into* the workload you
> > > care about...
> > 
> > Ah, that was only intentionally done for the purposes of measuring the
> > impact.  That's not done on purpose in production.
> > 
> > Nevertheless, the example we tend to run into is that a box may have 2
> > NICs and each NIC has 32 Tx-Rx queues.  This means we've got 64 NIC
> > interrupts, each assigned to a different CPU.  Our systems have 64 CPUs.
> > What happens in practice is that a VM will get torn down, and that has a
> > 1-in-64 chance of impacting the performance of the subset of the flows
> > that are mapped via RSS to the interrupt that happens to be assigned to
> > the CPU where the VM is being torn down.
> > 
> > Of course, the obvious next question is why not just bind the VMs flows
> > to the CPUs the VM is running on?  We don't have a 1:1 mapping of
> > network device to VM, or VM to CPU right now, which frustrates this
> > approach.
> > 
> > > > Further reducing the stage2_apply_range() batch size has substantial
> > > > performance improvements for IO that share a CPU performing an unmap
> > > > operation.  By switching to a 2mb chunk, IO performance regressions were
> > > > no longer observed in this author's tests.  E.g. it was possible to
> > > > obtain the advertised device throughput despite an unmap operation
> > > > occurring on the CPU where the interrupt was running.  There is a
> > > > tradeoff, however.  No changes were observed in per-operation timings
> > > > when running the kvm_pagetable_test without an interrupt load.  However,
> > > > with a 64gb VM, 1 vcpu, and 4k pages and a IO load, map times increased
> > > > by about 15% and unmap times increased by about 58%.  In essence, this
> > > > trades slower map/unmap times for improved IO throughput.
> > > 
> > > There are other users of the range-based operations, like
> > > write-protection. Live migration is especially sensitive to the latency
> > > of page table updates as it can affect the VMM's ability to converge
> > > with the guest.
> > 
> > To be clear, the reduction in performance was observed when I
> > concurrently executed both the kvm_pagetable_test and a networking
> > benchmark where the NIC's interrupts were assigned to the same CPU where
> > the pagetable test was executing.  I didn't see a slowdown just running
> > the pagetable test.
> 
> Any chance you could share more details about your HW configuration
> (what CPU is that?)  and the type of traffic? This is the sort of
> things I'd like to be able to reproduce in order to experiment various
> strategies.

Sure, I only have access to documentation that is publicly available.

The hardware where we ran into this inititally was Graviton 3, which is
a Neoverse-V1 based core.  It does not support FEAT_TLBIRANGE.  I've
also tested on Graviton 4, which is Neoverse-V2 based.  It _does_
support FEAT_TLBIRANGE.  The deferred range based invalidation
support, was enough to allow us to teardown a large VM based on 4k pages
and not incur a visible performance penalty.  I haven't had a chance to
test to see if and how Will's patches change this, though.

The tests themselves were not especially fancy. The networking hardware
was a ENA device on an EC2 box with 30Gbps limit (5/10 Gbps per flow,
depending on the config).  The storage tested was a gp3 EBS device
configured to max IOPS/throughput (16,000 IOPS / 1000Mb/s).

Networking tests were iperf3 with a 9001 byte packet size.  The storage
tests were fio's randwrite workload in directio mode using the libaio
backend.  The "IOPS" test used a 4k blocksize and a queue depth of 128.
The "throughput" test used a blocksize of 64k and an iodepth of 32.  For
the fio tests, it was a 10gb file and 2 workers, mostly because the EBS
devices have two hardware queues for data.

I ran the kvm_page_table_test with a few different sizes, but settled on
64G with 1 vcpu for most tests.

Let me know if there's anything else I can share here.

-K