[PATCH] irqchip: gicv3-its: Use NUMA aware memory allocation for ITS tables

[Jayachandran C]

Hi Marc,

On Mon, Jul 10, 2017 at 04:15:28PM +0100, Marc Zyngier wrote:
> On 10/07/17 15:57, Shanker Donthineni wrote:
> > Hi Marc,
> > 
> > On 07/10/2017 08:50 AM, Marc Zyngier wrote:
> >> On 10/07/17 11:21, Ganapatrao Kulkarni wrote:
> >>> Hi Marc,
> >>>
> >>> On Mon, Jul 10, 2017 at 2:53 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
> >>>> On 10/07/17 10:08, Ganapatrao Kulkarni wrote:
> >>>>> On Mon, Jul 10, 2017 at 2:36 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
> >>>>>> On 10/07/17 09:48, Ganapatrao Kulkarni wrote:
> >>>>>>> Hi Marc,
> >>>>>>>
> >>>>>>> On Mon, Jul 3, 2017 at 8:23 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
> >>>>>>>> Hi Shanker,
> >>>>>>>>
> >>>>>>>> On 03/07/17 15:24, Shanker Donthineni wrote:
> >>>>>>>>> Hi Marc,
> >>>>>>>>>
> >>>>>>>>> On 06/30/2017 03:51 AM, Marc Zyngier wrote:
> >>>>>>>>>> On 30/06/17 04:01, Ganapatrao Kulkarni wrote:
> >>>>>>>>>>> On Fri, Jun 30, 2017 at 8:04 AM, Ganapatrao Kulkarni
> >>>>>>>>>>> <gpkulkarni at gmail.com> wrote:
> >>>>>>>>>>>> Hi Shanker,
> >>>>>>>>>>>>
> >>>>>>>>>>>> On Sun, Jun 25, 2017 at 9:16 PM, Shanker Donthineni
> >>>>>>>>>>>> <shankerd at codeaurora.org> wrote:
> >>>>>>>>>>>>> The NUMA node information is visible to ITS driver but not being used
> >>>>>>>>>>>>> other than handling errata. This patch allocates the memory for ITS
> >>>>>>>>>>>>> tables from the corresponding NUMA node using the appropriate NUMA
> >>>>>>>>>>>>> aware functions.
> >>>>>>>>>>>
> >>>>>>>>>>> IMHO, the description would have been more constructive?
> >>>>>>>>>>>
> >>>>>>>>>>> "All ITS tables are mapped by default to NODE 0 memory.
> >>>>>>>>>>> Adding changes to allocate memory from respective NUMA NODES of ITS devices.
> >>>>>>>>>>> This will optimize tables access and avoids unnecessary inter-node traffic."
> >>>>>>>>>>
> >>>>>>>>>> But more importantly, I'd like to see figures showing the actual benefit
> >>>>>>>>>> of this per-node allocation. Given that both of you guys have access to
> >>>>>>>>>> such platforms, please show me the numbers!
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> I'll share the actual results which shows the improvement whenever
> >>>>>>>>> available on our next chips. Current version of Qualcomm qdf2400 doesn't
> >>>>>>>>> support multi socket configuration to capture results and share with you.
> >>>>>>>>>
> >>>>>>>>> Do you see any other issues with this patch apart from the performance
> >>>>>>>>> improvements. I strongly believe this brings the noticeable improvement
> >>>>>>>>> in numbers on systems where it has multi node memory/CPU configuration.
> >>>>>>>>
> >>>>>>>> I agree that it *could* show an improvement, but it very much depends on
> >>>>>>>> how often the ITS misses in its caches. For this kind of patches, I want
> >>>>>>>> to see two things:
> >>>>>>>>
> >>>>>>>> 1) It brings a measurable benefit on NUMA platforms
> >>>>>>>
> >>>>>>> Did some measurement of interrupt response time for LPIs and we don't
> >>>>>>> see any major
> >>>>>>> improvement due to caching of Tables. However, we have seen
> >>>>>>> improvements of around 5%.
> >>>>>>
> >>>>>> An improvement of what exactly?
> >>>>>
> >>>>> interrupt response time.
> >>>>
> >>>> Measured how? On which HW? Using which benchmark?
> >>>
> >>> This has been tested on ThunderX2.
> >>> We have instrumented gic-v3-its driver code to create dummy LPI device
> >>> with few vectors.
> >>> The LPI is induced from dummy device(through sysfs by writing to
> >>> TRANSLATOR reg).
> >>> The ISR routine(gic_handle_irq) being called to handle the induced LPI.
> >>> NODE 1 cpu is used to induce LPI and NODE 1 cpu/collection is mapped
> >>> in ITT to route this LPI.
> >>>
> >>> CPU timer counter are sampled at the time LPI is Induced and in ISR
> >>> routine to calculate interrupt response time.
> >>> the result shown improvement of 5% with this patch.
> >>
> >> And you call that a realistic measurement of the latency? Really? Sorry,
> >> but I cannot take you seriously here.
> >>
> >>> Do you have any recommended benchmarks to test the same?
> >>
> >> Run a standard benchmark such as netperf, post the result with and
> >> without that patch. The above is just plain ridiculous.
> >>
> > 
> > The whole purpose of ACPI subtable "GIC Interrupt Translation Service (ITS) Affinity structure" 
> > is to provide the proximity information to OS so that software will take advantage of NUMA
> > aware allocations to improve the read latency of ITS/GICR tables, not just for implementing 
> > software workarounds.
> > 
> >  
> > I believe ITS driver should provide NUMA aware allocations just like x86 Linux drivers. How much
> > performance improvement we observer is based on the individual SOC implementation, inter NODE
> > latency, inter node traffic, cache capacity, and type of the test used to measure results.
> > 
> > Please consider this patch irrespective of the test results running on a specific hardware. We
> > need this patch for upcoming Qualcomm server chips. 
> 
> "I believe" and "We need" are not a proof of the usefulness of this. We
> can argue all day, or you can provide a set of convincing results. Your
> choice. But I can guarantee you the the latter is a much better method
> than the former.
> 
> If you (or Cavium) cannot be bothered to provide tangible results that
> this is useful, why should I take this at face value? This is just like
> any other improvement we make to the kernel. We back it *with data*.

At Cavium, most of the ThunderX2 boards we have are multi-node, and we
are interested in enabling NUMA optimizations.

But, in this case, we do not see (or expect to see - given the nature of
access) any significant improvement in any standard benchmark. Ganapat's
LPI injection test to find interrupt latency was probably the best option
we had so far. We could come up with another contrived test case to see
if there is any change in behavior when we overload the interconnect,
but I don't think we will get any data to really justify the patch.

Allocating the tables on the node is a good thing since it avoids
unnecessary traffic over the interconnect, so I do not see the
problem in merging a simple patch for that. Is there any specific
issue here?

Anyway, for ThunderX2, the patch is good to have, but not critical.
And as Ganapat noted, the patch can be improved a bit. Also going thru
the patch, I think the chip data is better allocated using node as well.

JC.