Re: [RFC PATCH] KVM: arm/arm64: Enable direct irqfd MSI injection

[Raslan, KarimAllah]

On Mon, 2019-05-20 at 23:31 +0800, Zenghui Yu wrote:
> Hi Marc,
> 
> On 2019/5/16 15:21, Marc Zyngier wrote:
> > 
> > Hi Andre,
> > 
> > On Wed, 15 May 2019 17:38:32 +0100,
> > Andre Przywara <andre.przywara@xxxxxxx> wrote:
> > > 
> > > 
> > > On Mon, 18 Mar 2019 13:30:40 +0000
> > > Marc Zyngier <marc.zyngier@xxxxxxx> wrote:
> > > 
> > > Hi,
> > > 
> > > > 
> > > > On Sun, 17 Mar 2019 19:35:48 +0000
> > > > Marc Zyngier <marc.zyngier@xxxxxxx> wrote:
> > > > 
> > > > [...]
> > > > 
> > > > > 
> > > > > A first approach would be to keep a small cache of the last few
> > > > > successful translations for this ITS, cache that could be looked-up by
> > > > > holding a spinlock instead. A hit in this cache could directly be
> > > > > injected. Any command that invalidates or changes anything (DISCARD,
> > > > > INV, INVALL, MAPC with V=0, MAPD with V=0, MOVALL, MOVI) should nuke
> > > > > the cache altogether.
> > > > 
> > > > And to explain what I meant with this, I've pushed a branch[1] with a
> > > > basic prototype. It is good enough to get a VM to boot, but I wouldn't
> > > > trust it for anything serious just yet.
> > > > 
> > > > If anyone feels like giving it a go and check whether it has any
> > > > benefit performance wise, please do so.
> > > 
> > > So I took a stab at the performance aspect, and it took me a while to find
> > > something where it actually makes a difference. The trick is to create *a
> > > lot* of interrupts. This is my setup now:
> > > - GICv3 and ITS
> > > - 5.1.0 kernel vs. 5.1.0 plus Marc's rebased "ITS cache" patches on top
> > > - 4 VCPU guest on a 4 core machine
> > > - passing through a M.2 NVMe SSD (or a USB3 controller) to the guest
> > > - running FIO in the guest, with:
> > >    - 4K block size, random reads, queue depth 16, 4 jobs (small)
> > >    - 1M block size, sequential reads, QD 1, 1 job (big)
> > > 
> > > For the NVMe disk I see a whopping 19% performance improvement with Marc's
> > > series (for the small blocks). For a SATA SSD connected via USB3.0 I still
> > > see 6% improvement. For NVMe there were 50,000 interrupts per second on
> > > the host, the USB3 setup came only up to 10,000/s. For big blocks (with
> > > IRQs in the low thousands/s) the win is less, but still a measurable
> > > 3%.
> > 
> > Thanks for having a go at this, and identifying the case where it
> > actually matters (I would have hoped that the original reporter would
> > have helped with this, but hey, never mind). The results are pretty
> > impressive (more so than I anticipated), and I wonder whether we could
> > improve things further (50k interrupts/s is not that high -- I get
> > more than 100k on some machines just by playing with their sdcard...).
> 
> I think the "original reporter" must feel embarrassed now.
> Actually, we had tested your patches (based on about 5.1.0-rc2) but
> failed to see performance improvement. And I stopped to move on, and
> then two months had gone... Oh sorry!
> 
> We retest your patches on 5.1.0, the result is as below.
> 
> Test setup:
> - GICv3 and ITS (on Taishan 2280, D05)
> - two 4-VCPU guests with vhost-net interface
> - run iperf in guests:
>     - guest1: iperf -s
>     - guest2: iperf -c guest1-IP -t 10
> - pin vcpu threads and vhost threads on the same NUMA node
> 
> Result:
> +-----------------+--------------+-----------------------+
> > 
> > Result          | interrupts/s | bandwidth (Gbits/sec) |
> +-----------------+--------------+-----------------------+
> > 
> > 5.1.0           |    25+ k     |    10.6 Gbits/sec     |
> +-----------------+--------------+-----------------------+
> > 
> > 5.1.0 (patched) |    40+ k     |    10.2 Gbits/sec     |
> +-----------------+--------------+-----------------------+
> 
> We get "interrupts/s" from /proc/interrupts on iperf server, with stable
> measured results. And we get "bandwidth" directly from iperf, but the
> results are somewhat *instable*. And the results really confused me --
> we received more interrupts but get a slight lower performance, why?
> 
> We configure the vhost-net interface with only one queue, so I think we
> can rule out the spin-lock influence. And 'perf lock' confirmed this.
> This is all that I can provide now, sorry if it's useless.
> 
> Also, one minor nit in code:
> In vgic_its_cache_translation(), we use vgic_put_irq() to evict the LRU
> cache entry, while we're already holding the lpi_list_lock. A deadlock
> will be caused here. But this is easy to fix.
> 
> 
> Anyway, we always have enough environments (e.g., D05, D06, ...) to do
> some tests. If you want to do further tests on our boards, please let me
> know :)

You actually need a little bit more control over the interrupt pinning in theÂ
guest and the interrupt pinning on the host. You also need to control vCPUÂ
pinning on pCPUs to have deterministic benchmark here.

I have a patch (which is not as polished as the one from Marc) that does directÂ
interrupt injection and we do see roughly 20%-25% bandwidth increase. So yes,Â
direct interrupt injection is absolutely needed.

Generally, if you have the host interrupt hitting a CPU different from the onesÂ
running the guest you see higherÂbandwidth (with current vanilla KVM). Once theÂ
host interrupts hit the same vCPU as the guest, not having this direct injection
path shows a huge drop in bandwidth.

So generally I would suggest to move forward with direct injection patch as itÂ
is really needed in platforms that does not have "posted interrupts".

> 
> 
> thanks,
> zenghui
> 
> > 
> > Could you describe how many interrupt sources each device has? The
> > reason I'm asking is that the cache size is pretty much hardcoded at
> > the moment (4 entries per vcpu), and that could have an impact on
> > performance if we keep evicting entries in the cache (note to self:
> > add some statistics for that).
> > 
> > Another area where we can improve things is that I think the
> > invalidation mechanism is pretty trigger happy (MOVI really doesn't
> > need to invalidate the cache). On the other hand, I'm not sure your
> > guest does too much of that.
> > 
> > Finally, the single cache spin-lock is bound to be a bottleneck of its
> > own at high interrupt rates, and I wonder whether we should move the
> > whole thing over to an RCU friendly data structure (the vgic_irq
> > structure really isn't that friendly). It'd be good to find out how
> > contended that spinlock is on your system.
> > 
> > > 
> > > Now that I have the setup, I can rerun experiments very quickly (given I
> > > don't loose access to the machine), so let me know if someone needs
> > > further tests.
> > 
> > Another useful data point would be the delta with bare-metal: how much
> > overhead do we have with KVM, with and without this patch series. Oh,
> > and for easier comparison, please write it as a table that we can dump
> > in the cover letter when I actually post the series! ;-)
> > 
> > Thanks,
> > 
> > 	M.
> > 
> 



Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrer: Christian Schlaeger, Ralf Herbrich
Ust-ID: DE 289 237 879
Eingetragen am Amtsgericht Charlottenburg HRB 149173 B