| Message ID | 20211105105136.12137-1-21cnbao@gmail.com (mailing list archive) |
|---|---|
| State | RFC |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | [RFC] sched&net: avoid over-pulling tasks due to network interrupts | expand |
On Fri, Nov 05, 2021 at 06:51:36PM +0800, Barry Song wrote: > From: Barry Song <song.bao.hua@hisilicon.com> > > In LPC2021, both Libo Chen and Tim Chen have reported the overpull > of network interrupts[1]. For example, while running one database, > ethernet is located in numa0, numa1 might be almost idle due to > interrupts are pulling tasks to numa0 because of wake_up affine. > I have seen the same problem. One way to solve this problem is > moving to a normal wakeup in network rather than using a sync > wakeup which will be more aggressively pulling tasks in scheduler > core. > > On kunpeng920 with 4numa, ethernet is located at numa0, storage > disk is located at numa2. While using sysbench to connect this > mysql machine, I am seeing numa1 is idle though numa0,2 and 3 > are quite busy. > > I am not saying this patch is exactly the right approach, But I'd > like to use this RFC to connect the people of net and scheduler, > and start the discussion in this wider range. Well the normal way would be to use multi-queue crud and/or receive packet steering to get the interrupt/wakeup back to the cpu that data came from.
On Sat, Nov 6, 2021 at 1:25 AM Peter Zijlstra <peterz@infradead.org> wrote: > > On Fri, Nov 05, 2021 at 06:51:36PM +0800, Barry Song wrote: > > From: Barry Song <song.bao.hua@hisilicon.com> > > > > In LPC2021, both Libo Chen and Tim Chen have reported the overpull > > of network interrupts[1]. For example, while running one database, > > ethernet is located in numa0, numa1 might be almost idle due to > > interrupts are pulling tasks to numa0 because of wake_up affine. > > I have seen the same problem. One way to solve this problem is > > moving to a normal wakeup in network rather than using a sync > > wakeup which will be more aggressively pulling tasks in scheduler > > core. > > > > On kunpeng920 with 4numa, ethernet is located at numa0, storage > > disk is located at numa2. While using sysbench to connect this > > mysql machine, I am seeing numa1 is idle though numa0,2 and 3 > > are quite busy. > > > > > I am not saying this patch is exactly the right approach, But I'd > > like to use this RFC to connect the people of net and scheduler, > > and start the discussion in this wider range. > > Well the normal way would be to use multi-queue crud and/or receive > packet steering to get the interrupt/wakeup back to the cpu that data > came from. The test case has been a multi-queue ethernet and irqs are balanced to NUMA0 by irqbalanced or pinned to NUMA0 where the card is located by the script like: #!/bin/bash irq_list=(`cat /proc/interrupts | grep network_name| awk -F: '{print $1}'`) cpunum=0 for irq in ${irq_list[@]} do echo $cpunum > /proc/irq/$irq/smp_affinity_list echo `cat /proc/irq/$irq/smp_affinity_list` (( cpunum+=1 )) done I have heard some people are working around this issue by pinning multi-queue IRQs to multiple NUMAs which can spread interrupts and avoid over-pulling tasks to one NUMA only, but lose ethernet locality? Hi, @Tim, it seems in LPC2021 you mentioned you are using this solution? And some other people are pinning ethernet IRQs to a couple of CPUs within the NUMA ethernet belongs to, and then isolate these CPUs from tasks and use those CPUs for interrupts only. This can avoid wake-up pulling at all. I think we need some generic way to resolve this problem. Hi, @Libo , what is your solution to work around this issue? Thanks Barry
On Mon, Nov 08, 2021 at 07:08:09AM +1300, Barry Song wrote: > On Sat, Nov 6, 2021 at 1:25 AM Peter Zijlstra <peterz@infradead.org> wrote: > > > > On Fri, Nov 05, 2021 at 06:51:36PM +0800, Barry Song wrote: > > > From: Barry Song <song.bao.hua@hisilicon.com> > > > > > > In LPC2021, both Libo Chen and Tim Chen have reported the overpull > > > of network interrupts[1]. For example, while running one database, > > > ethernet is located in numa0, numa1 might be almost idle due to > > > interrupts are pulling tasks to numa0 because of wake_up affine. > > > I have seen the same problem. One way to solve this problem is > > > moving to a normal wakeup in network rather than using a sync > > > wakeup which will be more aggressively pulling tasks in scheduler > > > core. > > > > > > On kunpeng920 with 4numa, ethernet is located at numa0, storage > > > disk is located at numa2. While using sysbench to connect this > > > mysql machine, I am seeing numa1 is idle though numa0,2 and 3 > > > are quite busy. > > > > > > > > I am not saying this patch is exactly the right approach, But I'd > > > like to use this RFC to connect the people of net and scheduler, > > > and start the discussion in this wider range. > > > > Well the normal way would be to use multi-queue crud and/or receive > > packet steering to get the interrupt/wakeup back to the cpu that data > > came from. > > The test case has been a multi-queue ethernet and irqs are balanced > to NUMA0 by irqbalanced or pinned to NUMA0 where the card is located > by the script like: > #!/bin/bash > irq_list=(`cat /proc/interrupts | grep network_name| awk -F: '{print $1}'`) > cpunum=0 > for irq in ${irq_list[@]} > do > echo $cpunum > /proc/irq/$irq/smp_affinity_list > echo `cat /proc/irq/$irq/smp_affinity_list` > (( cpunum+=1 )) > done > > I have heard some people are working around this issue by pinning > multi-queue IRQs to multiple NUMAs which can spread interrupts and > avoid over-pulling tasks to one NUMA only, but lose ethernet locality? So you're doing explicitly the wrong thing with your script above and then complain the scheduler follows that and destroys your data locality? The network folks made RPS/RFS specifically to spread the processing of the packets back to the CPUs/Nodes the TX happened on to increase data locality. Why not use that?
On 11/8/21 1:27 AM, Peter Zijlstra wrote: > On Mon, Nov 08, 2021 at 07:08:09AM +1300, Barry Song wrote: >> On Sat, Nov 6, 2021 at 1:25 AM Peter Zijlstra <peterz@infradead.org> wrote: >>> >>> On Fri, Nov 05, 2021 at 06:51:36PM +0800, Barry Song wrote: >>>> From: Barry Song <song.bao.hua@hisilicon.com> >>>> >>>> In LPC2021, both Libo Chen and Tim Chen have reported the overpull >>>> of network interrupts[1]. For example, while running one database, >>>> ethernet is located in numa0, numa1 might be almost idle due to >>>> interrupts are pulling tasks to numa0 because of wake_up affine. >>>> I have seen the same problem. One way to solve this problem is >>>> moving to a normal wakeup in network rather than using a sync >>>> wakeup which will be more aggressively pulling tasks in scheduler >>>> core. >>>> >>>> On kunpeng920 with 4numa, ethernet is located at numa0, storage >>>> disk is located at numa2. While using sysbench to connect this >>>> mysql machine, I am seeing numa1 is idle though numa0,2 and 3 >>>> are quite busy. >>>> >>> >>>> I am not saying this patch is exactly the right approach, But I'd >>>> like to use this RFC to connect the people of net and scheduler, >>>> and start the discussion in this wider range. >>> >>> Well the normal way would be to use multi-queue crud and/or receive >>> packet steering to get the interrupt/wakeup back to the cpu that data >>> came from. >> >> The test case has been a multi-queue ethernet and irqs are balanced >> to NUMA0 by irqbalanced or pinned to NUMA0 where the card is located >> by the script like: >> #!/bin/bash >> irq_list=(`cat /proc/interrupts | grep network_name| awk -F: '{print $1}'`) >> cpunum=0 >> for irq in ${irq_list[@]} >> do >> echo $cpunum > /proc/irq/$irq/smp_affinity_list >> echo `cat /proc/irq/$irq/smp_affinity_list` >> (( cpunum+=1 )) >> done >> >> I have heard some people are working around this issue by pinning >> multi-queue IRQs to multiple NUMAs which can spread interrupts and >> avoid over-pulling tasks to one NUMA only, but lose ethernet locality? > > So you're doing explicitly the wrong thing with your script above and > then complain the scheduler follows that and destroys your data > locality? > > The network folks made RPS/RFS specifically to spread the processing of > the packets back to the CPUs/Nodes the TX happened on to increase data > locality. Why not use that? > +1 This documentation should describe how this can be done Documentation/networking/scaling.rst Hopefully it is not completely outdated.
diff --git a/net/core/sock.c b/net/core/sock.c index 9862eef..a346359 100644 --- a/net/core/sock.c +++ b/net/core/sock.c @@ -3133,7 +3133,7 @@ void sock_def_readable(struct sock *sk) rcu_read_lock(); wq = rcu_dereference(sk->sk_wq); if (skwq_has_sleeper(wq)) - wake_up_interruptible_sync_poll(&wq->wait, EPOLLIN | EPOLLPRI | + wake_up_interruptible_poll(&wq->wait, EPOLLIN | EPOLLPRI | EPOLLRDNORM | EPOLLRDBAND); sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); rcu_read_unlock(); @@ -3151,7 +3151,7 @@ static void sock_def_write_space(struct sock *sk) if ((refcount_read(&sk->sk_wmem_alloc) << 1) <= READ_ONCE(sk->sk_sndbuf)) { wq = rcu_dereference(sk->sk_wq); if (skwq_has_sleeper(wq)) - wake_up_interruptible_sync_poll(&wq->wait, EPOLLOUT | + wake_up_interruptible_poll(&wq->wait, EPOLLOUT | EPOLLWRNORM | EPOLLWRBAND); /* Should agree with poll, otherwise some programs break */