Date:   Mon, 4 Dec 2023 09:02:08 +0000
From:   Souradeep Chakrabarti <schakrabarti@...rosoft.com>
To:     Yury Norov <yury.norov@...il.com>,
        Souradeep Chakrabarti <schakrabarti@...ux.microsoft.com>
CC:     Jakub Kicinski <kuba@...nel.org>,
        KY Srinivasan <kys@...rosoft.com>,
        Haiyang Zhang <haiyangz@...rosoft.com>,
        "wei.liu@...nel.org" <wei.liu@...nel.org>,
        Dexuan Cui <decui@...rosoft.com>,
        "davem@...emloft.net" <davem@...emloft.net>,
        "edumazet@...gle.com" <edumazet@...gle.com>,
        "pabeni@...hat.com" <pabeni@...hat.com>,
        Long Li <longli@...rosoft.com>,
        "sharmaajay@...rosoft.com" <sharmaajay@...rosoft.com>,
        "leon@...nel.org" <leon@...nel.org>,
        "cai.huoqing@...ux.dev" <cai.huoqing@...ux.dev>,
        "ssengar@...ux.microsoft.com" <ssengar@...ux.microsoft.com>,
        "vkuznets@...hat.com" <vkuznets@...hat.com>,
        "tglx@...utronix.de" <tglx@...utronix.de>,
        "linux-hyperv@...r.kernel.org" <linux-hyperv@...r.kernel.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-rdma@...r.kernel.org" <linux-rdma@...r.kernel.org>,
        Paul Rosswurm <paulros@...rosoft.com>
Subject: RE: [EXTERNAL] Re: [PATCH V2 net-next] net: mana: Assigning IRQ
 affinity on HT cores



>-----Original Message-----
>From: Yury Norov <yury.norov@...il.com>
>Sent: Thursday, November 30, 2023 10:27 PM
>To: Souradeep Chakrabarti <schakrabarti@...ux.microsoft.com>
>Cc: Souradeep Chakrabarti <schakrabarti@...rosoft.com>; Jakub Kicinski
><kuba@...nel.org>; KY Srinivasan <kys@...rosoft.com>; Haiyang Zhang
><haiyangz@...rosoft.com>; wei.liu@...nel.org; Dexuan Cui
><decui@...rosoft.com>; davem@...emloft.net; edumazet@...gle.com;
>pabeni@...hat.com; Long Li <longli@...rosoft.com>;
>sharmaajay@...rosoft.com; leon@...nel.org; cai.huoqing@...ux.dev;
>ssengar@...ux.microsoft.com; vkuznets@...hat.com; tglx@...utronix.de;
>linux-hyperv@...r.kernel.org; netdev@...r.kernel.org; linux-
>kernel@...r.kernel.org; linux-rdma@...r.kernel.org; Paul Rosswurm
><paulros@...rosoft.com>
>Subject: Re: [EXTERNAL] Re: [PATCH V2 net-next] net: mana: Assigning IRQ
>affinity on HT cores
>
>[You don't often get email from yury.norov@...il.com. Learn why this is
>important at https://aka.ms/LearnAboutSenderIdentification ]
>
>On Thu, Nov 30, 2023 at 04:05:12AM -0800, Souradeep Chakrabarti wrote:
>> On Wed, Nov 29, 2023 at 06:16:17PM -0800, Yury Norov wrote:
>> > On Mon, Nov 27, 2023 at 09:36:38AM +0000, Souradeep Chakrabarti
>wrote:
>> > >
>> > >
>> > > >-----Original Message-----
>> > > >From: Jakub Kicinski <kuba@...nel.org>
>> > > >Sent: Wednesday, November 22, 2023 5:19 AM
>> > > >To: Souradeep Chakrabarti <schakrabarti@...ux.microsoft.com>
>> > > >Cc: KY Srinivasan <kys@...rosoft.com>; Haiyang Zhang
>> > > ><haiyangz@...rosoft.com>; wei.liu@...nel.org; Dexuan Cui
>> > > ><decui@...rosoft.com>; davem@...emloft.net;
>edumazet@...gle.com;
>> > > >pabeni@...hat.com; Long Li <longli@...rosoft.com>;
>> > > >sharmaajay@...rosoft.com; leon@...nel.org; cai.huoqing@...ux.dev;
>> > > >ssengar@...ux.microsoft.com; vkuznets@...hat.com;
>> > > >tglx@...utronix.de; linux- hyperv@...r.kernel.org;
>> > > >netdev@...r.kernel.org; linux-kernel@...r.kernel.org;
>> > > >linux-rdma@...r.kernel.org; Souradeep Chakrabarti
>> > > ><schakrabarti@...rosoft.com>; Paul Rosswurm
>> > > ><paulros@...rosoft.com>
>> > > >Subject: [EXTERNAL] Re: [PATCH V2 net-next] net: mana: Assigning
>> > > >IRQ affinity on HT cores
>> > > >
>> > > >On Tue, 21 Nov 2023 05:54:37 -0800 Souradeep Chakrabarti wrote:
>> > > >> Existing MANA design assigns IRQ to every CPUs, including
>> > > >> sibling hyper-threads in a core. This causes multiple IRQs to
>> > > >> work on same CPU and may reduce the network performance with RSS.
>> > > >>
>> > > >> Improve the performance by adhering the configuration for RSS,
>> > > >> which assigns IRQ on HT cores.
>> > > >
>> > > >Drivers should not have to carry 120 LoC for something as basic as
>spreading IRQs.
>> > > >Please take a look at include/linux/topology.h and if there's
>> > > >nothing that fits your needs there - add it. That way other drivers can
>reuse it.
>> > > Because of the current design idea, it is easier to keep things
>> > > inside the mana driver code here. As the idea of IRQ distribution here is :
>> > > 1)Loop through interrupts to assign CPU 2)Find non sibling online
>> > > CPU from local NUMA and assign the IRQs on them.
>> > > 3)If number of IRQs is more than number of non-sibling CPU in that
>> > > NUMA node, then assign on sibling CPU of that node.
>> > > 4)Keep doing it till all the online CPUs are used or no more IRQs.
>> > > 5)If all CPUs in that node are used, goto next NUMA node with CPU.
>> > > Keep doing 2 and 3.
>> > > 6) If all CPUs in all NUMA nodes are used, but still there are
>> > > IRQs then wrap over from first local NUMA node and continue doing
>> > > 2, 3 4 till all IRQs are assigned.
>> >
>> > Hi Souradeep,
>> >
>> > (Thanks Jakub for sharing this thread with me)
>> >
>> > If I understand your intention right, you can leverage the existing
>> > cpumask_local_spread().
>> >
>> > But I think I've got something better for you. The below series adds
>> > a for_each_numa_cpu() iterator, which may help you doing most of the
>> > job without messing with nodes internals.
>> >
>> > https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Flo
>> > re.kernel.org%2Fnetdev%2FZD3l6FBnUh9vTIGc%40yury-
>ThinkPad%2FT%2F&dat
>> >
>a=05%7C01%7Cschakrabarti%40microsoft.com%7C79dfb421db6f463627250
>8dbf
>> >
>1c5c19e%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C6383696
>04095521
>> >
>996%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2lu
>MzIiLCJB
>> >
>TiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000%7C%7C%7C&sdata=pDUpYWo3K7
>uoz2q50GQ
>> > 1UKuPF2PwFagiT5pwrXhQXPk%3D&reserved=0
>> >
>> Thanks Yur and Jakub. I was trying to find this patch, but unable to find it on
>that thread.
>> Also in net-next I am unable to find it. Can you please tell, if it has been
>committed?
>> If not can you please point me out the correct patch for this macro.
>> It will be really helpful.
>
>Try this branch. I just rebased it on top of bitmap-for-next, but didn't re-test.
>You may need to exclude the "sched: drop for_each_numa_hop_mask()" patch.
>
>https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithu
>b.com%2Fnorov%2Flinux%2Fcommits%2Ffor_each_numa_cpu&data=05%7C0
>1%7Cschakrabarti%40microsoft.com%7C79dfb421db6f4636272508dbf1c5c1
>9e%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C638369604095
>529277%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV
>2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C2000%7C%7C%7C&sdata=W
>wmd%2BvQS7YHIwFKyL9OLd8iYttJ4ZIqQyxU3Ex8UOkY%3D&reserved=0
>
>> > By using it, the pseudocode implementing your algorithm may look
>> > like this:
>> >
>> >         unsigned int cpu, hop;
>> >         unsigned int irq = 0;
>> >
>> > again:
>> >         cpu = get_cpu();
>> >         node = cpu_to_node(cpu);
>> >         cpumask_copy(cpus, cpu_online_mask);
>> >
>> >         for_each_numa_cpu(cpu, hop, node, cpus) {
>> >                 /* All siblings are the same for IRQ spreading purpose */
>> >                 irq_set_affinity_and_hint(irq,
>> > topology_sibling_cpumask());
>> >
>> >                 /* One IRQ per sibling group */
>> >                 cpumask_andnot(cpus, cpus,
>> > topology_sibling_cpumask());
>> >
>> >                 if (++irq == num_irqs)
>> >                         break;
>> >         }
>> >
>> >         if (irq < num_irqs)
>> >                 goto again;
>> >
>> > (Completely not tested, just an idea.)
>> >
>> I have done similar kind of change for our driver, but constraint here
>> is that total number of IRQs can be equal to the total number of
>> online CPUs, in some setup. It is either equal to the number of online CPUs or
>maximum 64 IRQs if online CPUs are more than that.
>
>Not sure I understand you. If you're talking about my proposal, there's
>seemingly no constraints on number of CPUs/IRQs.
>
>> So my proposed change is following:
>>
>> +static int irq_setup(int *irqs, int nvec, int start_numa_node) {
>> +       cpumask_var_t node_cpumask;
>> +       int i, cpu, err = 0;
>> +       unsigned int  next_node;
>> +       cpumask_t visited_cpus;
>> +       unsigned int start_node = start_numa_node;
>> +       i = 0;
>> +       if (!alloc_cpumask_var(&node_cpumask, GFP_KERNEL)) {
>> +               err = -ENOMEM;
>> +               goto free_mask;
>> +       }
>> +       cpumask_andnot(&visited_cpus, &visited_cpus, &visited_cpus);
>> +       start_node = 1;
>> +       for_each_next_node_with_cpus(start_node, next_node) {
>
>If your goal is to maximize locality, this doesn't seem to be correct.
>for_each_next_node_with_cpus() is based on next_node(), and so enumerates
>nodes in a numerically increasing order. On real machines, it's possible that
>numerically adjacent node is not the topologically nearest.
>
>To approach that, for every node kernel maintains a list of equally distant nodes
>grouped into hops. You may likely want to use for_each_numa_hop_mask
>iterator, which iterated over hops in increasing distance order, instead of
>NUMA node numbers.
>
>But I would like to see for_each_numa_cpu() finally merged as a simpler and
>nicer alternative.
>
>> +               cpumask_copy(node_cpumask, cpumask_of_node(next_node));
>> +               for_each_cpu(cpu, node_cpumask) {
>> +                       cpumask_andnot(node_cpumask, node_cpumask,
>> +                                      topology_sibling_cpumask(cpu));
>> +                       irq_set_affinity_and_hint(irqs[i], cpumask_of(cpu));
>> +                       if(++i == nvec)
>> +                               goto free_mask;
>> +                       cpumask_set_cpu(cpu, &visited_cpus);
>> +                       if (cpumask_empty(node_cpumask) &&
>cpumask_weight(&visited_cpus) <
>> +                           nr_cpus_node(next_node)) {
>> +                               cpumask_copy(node_cpumask,
>cpumask_of_node(next_node));
>> +                               cpumask_andnot(node_cpumask, node_cpumask,
>&visited_cpus);
>> +                               cpu = cpumask_first(node_cpumask);
>> +                       }
>> +               }
>> +               if (next_online_node(next_node) == MAX_NUMNODES)
>> +                       next_node = first_online_node;
>> +       }
>> +free_mask:
>> +       free_cpumask_var(node_cpumask);
>> +       return err;
>> +}
>>
>> I can definitely use the for_each_numa_cpu() instead of my proposed
>> for_each_next_node_with_cpus() macro here and that will make it cleaner.
>> Thanks for the suggestion.
>
>Sure.
>
>Can you please share performance measurements for a solution you'll finally
>choose? Would be interesting to compare different approaches.
I have compared spreading IRQs across numa  with IRQs spread inside local
NUMA, and the performance was 14 percent better for later.
I have shared the V4 patch with for_each_numa_hop_mask() macro.
>
>Thanks,
>Yury

Powered by blists - more mailing lists