* [Question] ixgbe：Mechanism of RSS
@ 2025-01-02 3:53 Haifeng Xu
2025-01-02 8:13 ` Eric Dumazet
0 siblings, 1 reply; 15+ messages in thread
From: Haifeng Xu @ 2025-01-02 3:53 UTC (permalink / raw)
To: Tony Nguyen, , Przemek Kitszel, , "David S. Miller",
, Eric Dumazet, , Jakub Kicinski, , Paolo Abeni
Cc: linux-kernel, netdev, intel-wired-lan
Hi masters,
We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
The RSS configuration can be seen as follow:
RX flow hash indirection table for eno5 with 63 RX ring(s):
0: 0 1 2 3 4 5 6 7
8: 8 9 10 11 12 13 14 15
16: 0 1 2 3 4 5 6 7
24: 8 9 10 11 12 13 14 15
32: 0 1 2 3 4 5 6 7
40: 8 9 10 11 12 13 14 15
48: 0 1 2 3 4 5 6 7
56: 8 9 10 11 12 13 14 15
64: 0 1 2 3 4 5 6 7
72: 8 9 10 11 12 13 14 15
80: 0 1 2 3 4 5 6 7
88: 8 9 10 11 12 13 14 15
96: 0 1 2 3 4 5 6 7
104: 8 9 10 11 12 13 14 15
112: 0 1 2 3 4 5 6 7
120: 8 9 10 11 12 13 14 15
The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 3:53 [Question] ixgbe：Mechanism of RSS Haifeng Xu
@ 2025-01-02 8:13 ` Eric Dumazet
2025-01-02 8:43 ` Haifeng Xu
0 siblings, 1 reply; 15+ messages in thread
From: Eric Dumazet @ 2025-01-02 8:13 UTC (permalink / raw)
To: Haifeng Xu
Cc: Tony Nguyen, , Przemek Kitszel, , David S. Miller,
, Jakub Kicinski, , Paolo Abeni, linux-kernel, netdev,
intel-wired-lan
On Thu, Jan 2, 2025 at 4:53 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>
> Hi masters,
>
> We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
> The RSS configuration can be seen as follow:
>
> RX flow hash indirection table for eno5 with 63 RX ring(s):
> 0: 0 1 2 3 4 5 6 7
> 8: 8 9 10 11 12 13 14 15
> 16: 0 1 2 3 4 5 6 7
> 24: 8 9 10 11 12 13 14 15
> 32: 0 1 2 3 4 5 6 7
> 40: 8 9 10 11 12 13 14 15
> 48: 0 1 2 3 4 5 6 7
> 56: 8 9 10 11 12 13 14 15
> 64: 0 1 2 3 4 5 6 7
> 72: 8 9 10 11 12 13 14 15
> 80: 0 1 2 3 4 5 6 7
> 88: 8 9 10 11 12 13 14 15
> 96: 0 1 2 3 4 5 6 7
> 104: 8 9 10 11 12 13 14 15
> 112: 0 1 2 3 4 5 6 7
> 120: 8 9 10 11 12 13 14 15
>
> The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
>
> In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
It really depends on which cpus are assigned to each IRQ.
Look at /proc/irq/{IRQ_NUM}/smp_affinity
Also you can have some details in Documentation/networking/scaling.rst
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 8:13 ` Eric Dumazet
@ 2025-01-02 8:43 ` Haifeng Xu
2025-01-02 10:34 ` Eric Dumazet
0 siblings, 1 reply; 15+ messages in thread
From: Haifeng Xu @ 2025-01-02 8:43 UTC (permalink / raw)
To: Eric Dumazet
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On 2025/1/2 16:13, Eric Dumazet wrote:
> On Thu, Jan 2, 2025 at 4:53 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>>
>> Hi masters,
>>
>> We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
>> The RSS configuration can be seen as follow:
>>
>> RX flow hash indirection table for eno5 with 63 RX ring(s):
>> 0: 0 1 2 3 4 5 6 7
>> 8: 8 9 10 11 12 13 14 15
>> 16: 0 1 2 3 4 5 6 7
>> 24: 8 9 10 11 12 13 14 15
>> 32: 0 1 2 3 4 5 6 7
>> 40: 8 9 10 11 12 13 14 15
>> 48: 0 1 2 3 4 5 6 7
>> 56: 8 9 10 11 12 13 14 15
>> 64: 0 1 2 3 4 5 6 7
>> 72: 8 9 10 11 12 13 14 15
>> 80: 0 1 2 3 4 5 6 7
>> 88: 8 9 10 11 12 13 14 15
>> 96: 0 1 2 3 4 5 6 7
>> 104: 8 9 10 11 12 13 14 15
>> 112: 0 1 2 3 4 5 6 7
>> 120: 8 9 10 11 12 13 14 15
>>
>> The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
>>
>> In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
>
> It really depends on which cpus are assigned to each IRQ.
>
Hi Eric,
Each irq was assigned to a single cpu, for exapmle:
irq cpu
117 0
118 1
......
179 62
All cpus trigger interrupts not only cpus 0~15.
It seems that the result is inconsistent with the RSS hash value.
Thanks!
> Look at /proc/irq/{IRQ_NUM}/smp_affinity
>
> Also you can have some details in Documentation/networking/scaling.rst
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 8:43 ` Haifeng Xu
@ 2025-01-02 10:34 ` Eric Dumazet
2025-01-02 11:23 ` Haifeng Xu
0 siblings, 1 reply; 15+ messages in thread
From: Eric Dumazet @ 2025-01-02 10:34 UTC (permalink / raw)
To: Haifeng Xu
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On Thu, Jan 2, 2025 at 9:43 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>
>
>
> On 2025/1/2 16:13, Eric Dumazet wrote:
> > On Thu, Jan 2, 2025 at 4:53 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
> >>
> >> Hi masters,
> >>
> >> We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
> >> The RSS configuration can be seen as follow:
> >>
> >> RX flow hash indirection table for eno5 with 63 RX ring(s):
> >> 0: 0 1 2 3 4 5 6 7
> >> 8: 8 9 10 11 12 13 14 15
> >> 16: 0 1 2 3 4 5 6 7
> >> 24: 8 9 10 11 12 13 14 15
> >> 32: 0 1 2 3 4 5 6 7
> >> 40: 8 9 10 11 12 13 14 15
> >> 48: 0 1 2 3 4 5 6 7
> >> 56: 8 9 10 11 12 13 14 15
> >> 64: 0 1 2 3 4 5 6 7
> >> 72: 8 9 10 11 12 13 14 15
> >> 80: 0 1 2 3 4 5 6 7
> >> 88: 8 9 10 11 12 13 14 15
> >> 96: 0 1 2 3 4 5 6 7
> >> 104: 8 9 10 11 12 13 14 15
> >> 112: 0 1 2 3 4 5 6 7
> >> 120: 8 9 10 11 12 13 14 15
> >>
> >> The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
> >>
> >> In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
> >
> > It really depends on which cpus are assigned to each IRQ.
> >
>
> Hi Eric,
>
> Each irq was assigned to a single cpu, for exapmle:
>
> irq cpu
>
> 117 0
> 118 1
>
> ......
>
> 179 62
>
> All cpus trigger interrupts not only cpus 0~15.
> It seems that the result is inconsistent with the RSS hash value.
>
>
I misread your report, I thought you had 16 receive queues.
Why don't you change "ethtool -L eno5 rx 16", instead of trying to
configure RSS manually ?
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 10:34 ` Eric Dumazet
@ 2025-01-02 11:23 ` Haifeng Xu
2025-01-02 11:46 ` Eric Dumazet
2025-01-02 16:01 ` Edward Cree
0 siblings, 2 replies; 15+ messages in thread
From: Haifeng Xu @ 2025-01-02 11:23 UTC (permalink / raw)
To: Eric Dumazet
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On 2025/1/2 18:34, Eric Dumazet wrote:
> On Thu, Jan 2, 2025 at 9:43 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>>
>>
>>
>> On 2025/1/2 16:13, Eric Dumazet wrote:
>>> On Thu, Jan 2, 2025 at 4:53 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>>>>
>>>> Hi masters,
>>>>
>>>> We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
>>>> The RSS configuration can be seen as follow:
>>>>
>>>> RX flow hash indirection table for eno5 with 63 RX ring(s):
>>>> 0: 0 1 2 3 4 5 6 7
>>>> 8: 8 9 10 11 12 13 14 15
>>>> 16: 0 1 2 3 4 5 6 7
>>>> 24: 8 9 10 11 12 13 14 15
>>>> 32: 0 1 2 3 4 5 6 7
>>>> 40: 8 9 10 11 12 13 14 15
>>>> 48: 0 1 2 3 4 5 6 7
>>>> 56: 8 9 10 11 12 13 14 15
>>>> 64: 0 1 2 3 4 5 6 7
>>>> 72: 8 9 10 11 12 13 14 15
>>>> 80: 0 1 2 3 4 5 6 7
>>>> 88: 8 9 10 11 12 13 14 15
>>>> 96: 0 1 2 3 4 5 6 7
>>>> 104: 8 9 10 11 12 13 14 15
>>>> 112: 0 1 2 3 4 5 6 7
>>>> 120: 8 9 10 11 12 13 14 15
>>>>
>>>> The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
>>>>
>>>> In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
>>>
>>> It really depends on which cpus are assigned to each IRQ.
>>>
>>
>> Hi Eric,
>>
>> Each irq was assigned to a single cpu, for exapmle:
>>
>> irq cpu
>>
>> 117 0
>> 118 1
>>
>> ......
>>
>> 179 62
>>
>> All cpus trigger interrupts not only cpus 0~15.
>> It seems that the result is inconsistent with the RSS hash value.
>>
>>
>
> I misread your report, I thought you had 16 receive queues.
>
> Why don't you change "ethtool -L eno5 rx 16", instead of trying to
> configure RSS manually ?
Hi Eric,
We want to make full use of cpu resources to receive packets. So
we enable 63 rx queues. But we found the rate of interrupt growth
on cpu 0~15 is faster than other cpus(almost twice). I don't know
whether it is related to RSS configuration. We didn't make any changes
on the RSS configration after the server is up.
FYI, on another server, we use Mellanox Technologies MT27800 NIC.
The rate of interrupt growth on cpu 0~63 seems have little gap.
It's RSS configration can be seen as follow:
RX flow hash indirection table for ens2f0np0 with 63 RX ring(s):
0: 0 1 2 3 4 5 6 7
8: 8 9 10 11 12 13 14 15
16: 16 17 18 19 20 21 22 23
24: 24 25 26 27 28 29 30 31
32: 32 33 34 35 36 37 38 39
40: 40 41 42 43 44 45 46 47
48: 48 49 50 51 52 53 54 55
56: 56 57 58 59 60 61 62 0
64: 1 2 3 4 5 6 7 8
72: 9 10 11 12 13 14 15 16
80: 17 18 19 20 21 22 23 24
88: 25 26 27 28 29 30 31 32
96: 33 34 35 36 37 38 39 40
104: 41 42 43 44 45 46 47 48
112: 49 50 51 52 53 54 55 56
120: 57 58 59 60 61 62 0 1
128: 2 3 4 5 6 7 8 9
136: 10 11 12 13 14 15 16 17
144: 18 19 20 21 22 23 24 25
152: 26 27 28 29 30 31 32 33
160: 34 35 36 37 38 39 40 41
168: 42 43 44 45 46 47 48 49
176: 50 51 52 53 54 55 56 57
184: 58 59 60 61 62 0 1 2
192: 3 4 5 6 7 8 9 10
200: 11 12 13 14 15 16 17 18
208: 19 20 21 22 23 24 25 26
216: 27 28 29 30 31 32 33 34
224: 35 36 37 38 39 40 41 42
232: 43 44 45 46 47 48 49 50
240: 51 52 53 54 55 56 57 58
248: 59 60 61 62 0 1 2 3
I am confused that why ixgbe NIC can dispatch the packets
to the rx queues that not specified in RSS configuration.
Thanks!
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 11:23 ` Haifeng Xu
@ 2025-01-02 11:46 ` Eric Dumazet
2025-01-03 2:36 ` Haifeng Xu
2025-01-02 16:01 ` Edward Cree
1 sibling, 1 reply; 15+ messages in thread
From: Eric Dumazet @ 2025-01-02 11:46 UTC (permalink / raw)
To: Haifeng Xu
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On Thu, Jan 2, 2025 at 12:23 PM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>
>
>
> On 2025/1/2 18:34, Eric Dumazet wrote:
> > On Thu, Jan 2, 2025 at 9:43 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
> >>
> >>
> >>
> >> On 2025/1/2 16:13, Eric Dumazet wrote:
> >>> On Thu, Jan 2, 2025 at 4:53 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
> >>>>
> >>>> Hi masters,
> >>>>
> >>>> We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
> >>>> The RSS configuration can be seen as follow:
> >>>>
> >>>> RX flow hash indirection table for eno5 with 63 RX ring(s):
> >>>> 0: 0 1 2 3 4 5 6 7
> >>>> 8: 8 9 10 11 12 13 14 15
> >>>> 16: 0 1 2 3 4 5 6 7
> >>>> 24: 8 9 10 11 12 13 14 15
> >>>> 32: 0 1 2 3 4 5 6 7
> >>>> 40: 8 9 10 11 12 13 14 15
> >>>> 48: 0 1 2 3 4 5 6 7
> >>>> 56: 8 9 10 11 12 13 14 15
> >>>> 64: 0 1 2 3 4 5 6 7
> >>>> 72: 8 9 10 11 12 13 14 15
> >>>> 80: 0 1 2 3 4 5 6 7
> >>>> 88: 8 9 10 11 12 13 14 15
> >>>> 96: 0 1 2 3 4 5 6 7
> >>>> 104: 8 9 10 11 12 13 14 15
> >>>> 112: 0 1 2 3 4 5 6 7
> >>>> 120: 8 9 10 11 12 13 14 15
> >>>>
> >>>> The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
> >>>>
> >>>> In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
> >>>
> >>> It really depends on which cpus are assigned to each IRQ.
> >>>
> >>
> >> Hi Eric,
> >>
> >> Each irq was assigned to a single cpu, for exapmle:
> >>
> >> irq cpu
> >>
> >> 117 0
> >> 118 1
> >>
> >> ......
> >>
> >> 179 62
> >>
> >> All cpus trigger interrupts not only cpus 0~15.
> >> It seems that the result is inconsistent with the RSS hash value.
> >>
> >>
> >
> > I misread your report, I thought you had 16 receive queues.
> >
> > Why don't you change "ethtool -L eno5 rx 16", instead of trying to
> > configure RSS manually ?
>
> Hi Eric,
>
> We want to make full use of cpu resources to receive packets. So
> we enable 63 rx queues. But we found the rate of interrupt growth
> on cpu 0~15 is faster than other cpus(almost twice). I don't know
> whether it is related to RSS configuration. We didn't make any changes
> on the RSS configration after the server is up.
>
>
>
> FYI, on another server, we use Mellanox Technologies MT27800 NIC.
> The rate of interrupt growth on cpu 0~63 seems have little gap.
>
> It's RSS configration can be seen as follow:
>
> RX flow hash indirection table for ens2f0np0 with 63 RX ring(s):
> 0: 0 1 2 3 4 5 6 7
> 8: 8 9 10 11 12 13 14 15
> 16: 16 17 18 19 20 21 22 23
> 24: 24 25 26 27 28 29 30 31
> 32: 32 33 34 35 36 37 38 39
> 40: 40 41 42 43 44 45 46 47
> 48: 48 49 50 51 52 53 54 55
> 56: 56 57 58 59 60 61 62 0
> 64: 1 2 3 4 5 6 7 8
> 72: 9 10 11 12 13 14 15 16
> 80: 17 18 19 20 21 22 23 24
> 88: 25 26 27 28 29 30 31 32
> 96: 33 34 35 36 37 38 39 40
> 104: 41 42 43 44 45 46 47 48
> 112: 49 50 51 52 53 54 55 56
> 120: 57 58 59 60 61 62 0 1
> 128: 2 3 4 5 6 7 8 9
> 136: 10 11 12 13 14 15 16 17
> 144: 18 19 20 21 22 23 24 25
> 152: 26 27 28 29 30 31 32 33
> 160: 34 35 36 37 38 39 40 41
> 168: 42 43 44 45 46 47 48 49
> 176: 50 51 52 53 54 55 56 57
> 184: 58 59 60 61 62 0 1 2
> 192: 3 4 5 6 7 8 9 10
> 200: 11 12 13 14 15 16 17 18
> 208: 19 20 21 22 23 24 25 26
> 216: 27 28 29 30 31 32 33 34
> 224: 35 36 37 38 39 40 41 42
> 232: 43 44 45 46 47 48 49 50
> 240: 51 52 53 54 55 56 57 58
> 248: 59 60 61 62 0 1 2 3
>
>
> I am confused that why ixgbe NIC can dispatch the packets
> to the rx queues that not specified in RSS configuration.
Perhaps make sure to change RX flow hash indirection table on the
Intel NIC then...
Maybe you changed the default configuration.
ethtool -X eno5 equal 64
Or
ethtool -X eno5 default
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 11:23 ` Haifeng Xu
2025-01-02 11:46 ` Eric Dumazet
@ 2025-01-02 16:01 ` Edward Cree
2025-01-02 16:39 ` Jakub Kicinski
2025-01-03 3:05 ` Haifeng Xu
1 sibling, 2 replies; 15+ messages in thread
From: Edward Cree @ 2025-01-02 16:01 UTC (permalink / raw)
To: Haifeng Xu, Eric Dumazet
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On 02/01/2025 11:23, Haifeng Xu wrote:
> We want to make full use of cpu resources to receive packets. So
> we enable 63 rx queues. But we found the rate of interrupt growth
> on cpu 0~15 is faster than other cpus(almost twice).
...
> I am confused that why ixgbe NIC can dispatch the packets
> to the rx queues that not specified in RSS configuration.
Hypothesis: it isn't doing so, RX is only happening on cpus (and
queues) 0-15, but the other CPUs are still sending traffic and
thus getting TX completion interrupts from their TX queues.
`ethtool -S` output has per-queue traffic stats which should
confirm this.
(But Eric is right that if you _want_ RX to use every CPU you
should just change the indirection table.)
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 16:01 ` Edward Cree
@ 2025-01-02 16:39 ` Jakub Kicinski
2025-01-03 2:37 ` Haifeng Xu
2025-01-03 3:05 ` Haifeng Xu
1 sibling, 1 reply; 15+ messages in thread
From: Jakub Kicinski @ 2025-01-02 16:39 UTC (permalink / raw)
To: Edward Cree
Cc: Haifeng Xu, Eric Dumazet, Tony Nguyen, Przemek Kitszel,
David S. Miller, Paolo Abeni, linux-kernel, netdev,
intel-wired-lan
On Thu, 2 Jan 2025 16:01:18 +0000 Edward Cree wrote:
> On 02/01/2025 11:23, Haifeng Xu wrote:
> > We want to make full use of cpu resources to receive packets. So
> > we enable 63 rx queues. But we found the rate of interrupt growth
> > on cpu 0~15 is faster than other cpus(almost twice).
> ...
> > I am confused that why ixgbe NIC can dispatch the packets
> > to the rx queues that not specified in RSS configuration.
>
> Hypothesis: it isn't doing so, RX is only happening on cpus (and
> queues) 0-15, but the other CPUs are still sending traffic and
> thus getting TX completion interrupts from their TX queues.
> `ethtool -S` output has per-queue traffic stats which should
> confirm this.
>
> (But Eric is right that if you _want_ RX to use every CPU you
> should just change the indirection table.)
IIRC Niantic had 4 bit entries in the RSS table or some such.
It wasn't possible to RSS across more than 16 queues at a time.
It's a great NIC but a bit dated at this point.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 11:46 ` Eric Dumazet
@ 2025-01-03 2:36 ` Haifeng Xu
0 siblings, 0 replies; 15+ messages in thread
From: Haifeng Xu @ 2025-01-03 2:36 UTC (permalink / raw)
To: Eric Dumazet
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On 2025/1/2 19:46, Eric Dumazet wrote:
> On Thu, Jan 2, 2025 at 12:23 PM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>>
>>
>>
>> On 2025/1/2 18:34, Eric Dumazet wrote:
>>> On Thu, Jan 2, 2025 at 9:43 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>>>>
>>>>
>>>>
>>>> On 2025/1/2 16:13, Eric Dumazet wrote:
>>>>> On Thu, Jan 2, 2025 at 4:53 AM Haifeng Xu <haifeng.xu@shopee.com> wrote:
>>>>>>
>>>>>> Hi masters,
>>>>>>
>>>>>> We use the Intel Corporation 82599ES NIC in our production environment. And it has 63 rx queues, every rx queue interrupt is processed by a single cpu.
>>>>>> The RSS configuration can be seen as follow:
>>>>>>
>>>>>> RX flow hash indirection table for eno5 with 63 RX ring(s):
>>>>>> 0: 0 1 2 3 4 5 6 7
>>>>>> 8: 8 9 10 11 12 13 14 15
>>>>>> 16: 0 1 2 3 4 5 6 7
>>>>>> 24: 8 9 10 11 12 13 14 15
>>>>>> 32: 0 1 2 3 4 5 6 7
>>>>>> 40: 8 9 10 11 12 13 14 15
>>>>>> 48: 0 1 2 3 4 5 6 7
>>>>>> 56: 8 9 10 11 12 13 14 15
>>>>>> 64: 0 1 2 3 4 5 6 7
>>>>>> 72: 8 9 10 11 12 13 14 15
>>>>>> 80: 0 1 2 3 4 5 6 7
>>>>>> 88: 8 9 10 11 12 13 14 15
>>>>>> 96: 0 1 2 3 4 5 6 7
>>>>>> 104: 8 9 10 11 12 13 14 15
>>>>>> 112: 0 1 2 3 4 5 6 7
>>>>>> 120: 8 9 10 11 12 13 14 15
>>>>>>
>>>>>> The maximum number of RSS queues is 16. So I have some questions about this. Will other cpus except 0~15 receive the rx interrupts?
>>>>>>
>>>>>> In our production environment, cpu 16~62 also receive the rx interrupts. Was our RSS misconfigured?
>>>>>
>>>>> It really depends on which cpus are assigned to each IRQ.
>>>>>
>>>>
>>>> Hi Eric,
>>>>
>>>> Each irq was assigned to a single cpu, for exapmle:
>>>>
>>>> irq cpu
>>>>
>>>> 117 0
>>>> 118 1
>>>>
>>>> ......
>>>>
>>>> 179 62
>>>>
>>>> All cpus trigger interrupts not only cpus 0~15.
>>>> It seems that the result is inconsistent with the RSS hash value.
>>>>
>>>>
>>>
>>> I misread your report, I thought you had 16 receive queues.
>>>
>>> Why don't you change "ethtool -L eno5 rx 16", instead of trying to
>>> configure RSS manually ?
>>
>> Hi Eric,
>>
>> We want to make full use of cpu resources to receive packets. So
>> we enable 63 rx queues. But we found the rate of interrupt growth
>> on cpu 0~15 is faster than other cpus(almost twice). I don't know
>> whether it is related to RSS configuration. We didn't make any changes
>> on the RSS configration after the server is up.
>>
>>
>>
>> FYI, on another server, we use Mellanox Technologies MT27800 NIC.
>> The rate of interrupt growth on cpu 0~63 seems have little gap.
>>
>> It's RSS configration can be seen as follow:
>>
>> RX flow hash indirection table for ens2f0np0 with 63 RX ring(s):
>> 0: 0 1 2 3 4 5 6 7
>> 8: 8 9 10 11 12 13 14 15
>> 16: 16 17 18 19 20 21 22 23
>> 24: 24 25 26 27 28 29 30 31
>> 32: 32 33 34 35 36 37 38 39
>> 40: 40 41 42 43 44 45 46 47
>> 48: 48 49 50 51 52 53 54 55
>> 56: 56 57 58 59 60 61 62 0
>> 64: 1 2 3 4 5 6 7 8
>> 72: 9 10 11 12 13 14 15 16
>> 80: 17 18 19 20 21 22 23 24
>> 88: 25 26 27 28 29 30 31 32
>> 96: 33 34 35 36 37 38 39 40
>> 104: 41 42 43 44 45 46 47 48
>> 112: 49 50 51 52 53 54 55 56
>> 120: 57 58 59 60 61 62 0 1
>> 128: 2 3 4 5 6 7 8 9
>> 136: 10 11 12 13 14 15 16 17
>> 144: 18 19 20 21 22 23 24 25
>> 152: 26 27 28 29 30 31 32 33
>> 160: 34 35 36 37 38 39 40 41
>> 168: 42 43 44 45 46 47 48 49
>> 176: 50 51 52 53 54 55 56 57
>> 184: 58 59 60 61 62 0 1 2
>> 192: 3 4 5 6 7 8 9 10
>> 200: 11 12 13 14 15 16 17 18
>> 208: 19 20 21 22 23 24 25 26
>> 216: 27 28 29 30 31 32 33 34
>> 224: 35 36 37 38 39 40 41 42
>> 232: 43 44 45 46 47 48 49 50
>> 240: 51 52 53 54 55 56 57 58
>> 248: 59 60 61 62 0 1 2 3
>>
>>
>> I am confused that why ixgbe NIC can dispatch the packets
>> to the rx queues that not specified in RSS configuration.
>
> Perhaps make sure to change RX flow hash indirection table on the
> Intel NIC then...
>
> Maybe you changed the default configuration.
>
> ethtool -X eno5 equal 64
The maximum number of RSS queues supported by Intel Corporation 82599ES NIC
is 16. When I specify the number which is larger than 16, it shows the below message.
"Cannot set RX flow hash configuration: Invalid argument."
> Or
> ethtool -X eno5 default
This command can run sucessfully and as I saied above, it only has 16 queues.
RX flow hash indirection table for eno5 with 63 RX ring(s):
0: 0 1 2 3 4 5 6 7
8: 8 9 10 11 12 13 14 15
16: 0 1 2 3 4 5 6 7
24: 8 9 10 11 12 13 14 15
32: 0 1 2 3 4 5 6 7
40: 8 9 10 11 12 13 14 15
48: 0 1 2 3 4 5 6 7
56: 8 9 10 11 12 13 14 15
64: 0 1 2 3 4 5 6 7
72: 8 9 10 11 12 13 14 15
80: 0 1 2 3 4 5 6 7
88: 8 9 10 11 12 13 14 15
96: 0 1 2 3 4 5 6 7
104: 8 9 10 11 12 13 14 15
112: 0 1 2 3 4 5 6 7
120: 8 9 10 11 12 13 14 15
Thanks!
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 16:39 ` Jakub Kicinski
@ 2025-01-03 2:37 ` Haifeng Xu
0 siblings, 0 replies; 15+ messages in thread
From: Haifeng Xu @ 2025-01-03 2:37 UTC (permalink / raw)
To: Jakub Kicinski, Edward Cree
Cc: Eric Dumazet, Tony Nguyen, Przemek Kitszel, David S. Miller,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On 2025/1/3 00:39, Jakub Kicinski wrote:
> On Thu, 2 Jan 2025 16:01:18 +0000 Edward Cree wrote:
>> On 02/01/2025 11:23, Haifeng Xu wrote:
>>> We want to make full use of cpu resources to receive packets. So
>>> we enable 63 rx queues. But we found the rate of interrupt growth
>>> on cpu 0~15 is faster than other cpus(almost twice).
>> ...
>>> I am confused that why ixgbe NIC can dispatch the packets
>>> to the rx queues that not specified in RSS configuration.
>>
>> Hypothesis: it isn't doing so, RX is only happening on cpus (and
>> queues) 0-15, but the other CPUs are still sending traffic and
>> thus getting TX completion interrupts from their TX queues.
>> `ethtool -S` output has per-queue traffic stats which should
>> confirm this.
>>
>> (But Eric is right that if you _want_ RX to use every CPU you
>> should just change the indirection table.)
>
> IIRC Niantic had 4 bit entries in the RSS table or some such.
> It wasn't possible to RSS across more than 16 queues at a time.
> It's a great NIC but a bit dated at this point.
Yes, It only has 16 RSS queues.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-02 16:01 ` Edward Cree
2025-01-02 16:39 ` Jakub Kicinski
@ 2025-01-03 3:05 ` Haifeng Xu
2025-01-07 17:16 ` Tony Nguyen
1 sibling, 1 reply; 15+ messages in thread
From: Haifeng Xu @ 2025-01-03 3:05 UTC (permalink / raw)
To: Edward Cree, Eric Dumazet
Cc: Tony Nguyen, Przemek Kitszel, David S. Miller, Jakub Kicinski,
Paolo Abeni, linux-kernel, netdev, intel-wired-lan
On 2025/1/3 00:01, Edward Cree wrote:
> On 02/01/2025 11:23, Haifeng Xu wrote:
>> We want to make full use of cpu resources to receive packets. So
>> we enable 63 rx queues. But we found the rate of interrupt growth
>> on cpu 0~15 is faster than other cpus(almost twice).
> ...
>> I am confused that why ixgbe NIC can dispatch the packets
>> to the rx queues that not specified in RSS configuration.
>
> Hypothesis: it isn't doing so, RX is only happening on cpus (and
> queues) 0-15, but the other CPUs are still sending traffic and
> thus getting TX completion interrupts from their TX queues.
> `ethtool -S` output has per-queue traffic stats which should
> confirm this.
>
I use ethtool -S to check the rx_queus stats and here is the result.
According to the below stats, all cpus have new packets received.
cpu t1(bytes) t2(bytes) delta(bytes)
0 154155550267550 154156433828875 883561325
1 148748566285840 148749509346247 943060407
2 148874911191685 148875798038140 886846455
3 152483460327704 152484251468998 791141294
4 147790981836915 147791775847804 794010889
5 146047892285722 146048778285682 885999960
6 142880516825921 142881213804363 696978442
7 152016735168735 152017707542774 972374039
8 146019936404393 146020739070311 802665918
9 147448522715540 147449258018186 735302646
10 145865736299432 145866601503106 865203674
11 149548527982122 149549289026453 761044331
12 146848384328236 146849303547769 919219533
13 152942139118542 152942769029253 629910711
14 150884661854828 150885556866976 895012148
15 149222733506734 149223510491115 776984381
16 34150226069524 34150375855113 149785589
17 34115700500819 34115914271025 213770206
18 33906215129998 33906448044501 232914503
19 33983812095357 33983986258546 174163189
20 34156349675011 34156565159083 215484072
21 33574293379024 33574490695725 197316701
22 33438129453422 33438297911151 168457729
23 32967454521585 32967612494711 157973126
24 33507443427266 33507604828468 161401202
25 33413275870121 33413433901940 158031819
26 33852322542796 33852527061150 204518354
27 33131162685385 33131330621474 167936089
28 33407661780251 33407823112381 161332130
29 34256799173845 34256944837757 145663912
30 33814458585183 33814623673528 165088345
31 33848638714862 33848775218038 136503176
32 18683932398308 18684069540891 137142583
33 19454524281229 19454647908293 123627064
34 19717744365436 19717900618222 156252786
35 20295086765202 20295245869666 159104464
36 20501853066588 20502000738936 147672348
37 20954631043374 20954797204375 166161001
38 21102911073326 21103062510369 151437043
39 21376404644179 21376515307288 110663109
40 20935812784743 20935983891491 171106748
41 20721278456831 20721435955715 157498884
42 21268291801465 21268425244578 133443113
43 21661413672829 21661629019091 215346262
44 21696437732484 21696568800049 131067565
45 21027869000890 21028020401214 151400324
46 21707137252644 21707293761990 156509346
47 20655623913790 20655740452889 116539099
48 32692002128477 32692138244468 136115991
49 33548445851486 33548569927672 124076186
50 33197264968787 33197448645817 183677030
51 33379544010500 33379746565576 202555076
52 33503579011721 33503722596159 143584438
53 33145734550468 33145892305819 157755351
54 33422692741858 33422844156764 151414906
55 32750945531107 32751131302251 185771144
56 33404955373530 33405157766253 202392723
57 33701185654471 33701313174725 127520254
58 33014531699810 33014700058409 168358599
59 32948906758429 32949151147605 244389176
60 33470813725985 33470993164755 179438770
61 33803771479735 33803971758441 200278706
62 33509751180818 33509926649969 175469151
Thanks!
> (But Eric is right that if you _want_ RX to use every CPU you
> should just change the indirection table.)
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-03 3:05 ` Haifeng Xu
@ 2025-01-07 17:16 ` Tony Nguyen
2025-01-08 3:36 ` Haifeng Xu
0 siblings, 1 reply; 15+ messages in thread
From: Tony Nguyen @ 2025-01-07 17:16 UTC (permalink / raw)
To: Haifeng Xu, Edward Cree, Eric Dumazet, Aleksandr Loktionov,
Kwapulinski, Piotr
Cc: Przemek Kitszel, David S. Miller, Jakub Kicinski, Paolo Abeni,
linux-kernel, netdev, intel-wired-lan
On 1/2/2025 7:05 PM, Haifeng Xu wrote:
>
>
> On 2025/1/3 00:01, Edward Cree wrote:
>> On 02/01/2025 11:23, Haifeng Xu wrote:
>>> We want to make full use of cpu resources to receive packets. So
>>> we enable 63 rx queues. But we found the rate of interrupt growth
>>> on cpu 0~15 is faster than other cpus(almost twice).
>> ...
>>> I am confused that why ixgbe NIC can dispatch the packets
>>> to the rx queues that not specified in RSS configuration.
>>
>> Hypothesis: it isn't doing so, RX is only happening on cpus (and
>> queues) 0-15, but the other CPUs are still sending traffic and
>> thus getting TX completion interrupts from their TX queues.
>> `ethtool -S` output has per-queue traffic stats which should
>> confirm this.
>>
>
> I use ethtool -S to check the rx_queus stats and here is the result.
>
> According to the below stats, all cpus have new packets received.
+ Alex and Piotr
What's your ntuple filter setting? If it's off, I suspect it may be the
Flow Director ATR (Application Targeting Routing) feature which will
utilize all queues. I believe if you turn on ntuple filters this will
turn that feature off.
Thanks,
Tony
>
> cpu t1(bytes) t2(bytes) delta(bytes)
>
> 0 154155550267550 154156433828875 883561325
> 1 148748566285840 148749509346247 943060407
> 2 148874911191685 148875798038140 886846455
> 3 152483460327704 152484251468998 791141294
> 4 147790981836915 147791775847804 794010889
> 5 146047892285722 146048778285682 885999960
> 6 142880516825921 142881213804363 696978442
> 7 152016735168735 152017707542774 972374039
> 8 146019936404393 146020739070311 802665918
> 9 147448522715540 147449258018186 735302646
> 10 145865736299432 145866601503106 865203674
> 11 149548527982122 149549289026453 761044331
> 12 146848384328236 146849303547769 919219533
> 13 152942139118542 152942769029253 629910711
> 14 150884661854828 150885556866976 895012148
> 15 149222733506734 149223510491115 776984381
> 16 34150226069524 34150375855113 149785589
> 17 34115700500819 34115914271025 213770206
> 18 33906215129998 33906448044501 232914503
> 19 33983812095357 33983986258546 174163189
> 20 34156349675011 34156565159083 215484072
> 21 33574293379024 33574490695725 197316701
> 22 33438129453422 33438297911151 168457729
> 23 32967454521585 32967612494711 157973126
> 24 33507443427266 33507604828468 161401202
> 25 33413275870121 33413433901940 158031819
> 26 33852322542796 33852527061150 204518354
> 27 33131162685385 33131330621474 167936089
> 28 33407661780251 33407823112381 161332130
> 29 34256799173845 34256944837757 145663912
> 30 33814458585183 33814623673528 165088345
> 31 33848638714862 33848775218038 136503176
> 32 18683932398308 18684069540891 137142583
> 33 19454524281229 19454647908293 123627064
> 34 19717744365436 19717900618222 156252786
> 35 20295086765202 20295245869666 159104464
> 36 20501853066588 20502000738936 147672348
> 37 20954631043374 20954797204375 166161001
> 38 21102911073326 21103062510369 151437043
> 39 21376404644179 21376515307288 110663109
> 40 20935812784743 20935983891491 171106748
> 41 20721278456831 20721435955715 157498884
> 42 21268291801465 21268425244578 133443113
> 43 21661413672829 21661629019091 215346262
> 44 21696437732484 21696568800049 131067565
> 45 21027869000890 21028020401214 151400324
> 46 21707137252644 21707293761990 156509346
> 47 20655623913790 20655740452889 116539099
> 48 32692002128477 32692138244468 136115991
> 49 33548445851486 33548569927672 124076186
> 50 33197264968787 33197448645817 183677030
> 51 33379544010500 33379746565576 202555076
> 52 33503579011721 33503722596159 143584438
> 53 33145734550468 33145892305819 157755351
> 54 33422692741858 33422844156764 151414906
> 55 32750945531107 32751131302251 185771144
> 56 33404955373530 33405157766253 202392723
> 57 33701185654471 33701313174725 127520254
> 58 33014531699810 33014700058409 168358599
> 59 32948906758429 32949151147605 244389176
> 60 33470813725985 33470993164755 179438770
> 61 33803771479735 33803971758441 200278706
> 62 33509751180818 33509926649969 175469151
>
> Thanks!
>
>> (But Eric is right that if you _want_ RX to use every CPU you
>> should just change the indirection table.)
>
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-07 17:16 ` Tony Nguyen
@ 2025-01-08 3:36 ` Haifeng Xu
2025-01-08 21:06 ` Tony Nguyen
0 siblings, 1 reply; 15+ messages in thread
From: Haifeng Xu @ 2025-01-08 3:36 UTC (permalink / raw)
To: Tony Nguyen, Edward Cree, Eric Dumazet, Aleksandr Loktionov,
Kwapulinski, Piotr
Cc: Przemek Kitszel, David S. Miller, Jakub Kicinski, Paolo Abeni,
linux-kernel, netdev, intel-wired-lan
On 2025/1/8 01:16, Tony Nguyen wrote:
>
>
> On 1/2/2025 7:05 PM, Haifeng Xu wrote:
>>
>>
>> On 2025/1/3 00:01, Edward Cree wrote:
>>> On 02/01/2025 11:23, Haifeng Xu wrote:
>>>> We want to make full use of cpu resources to receive packets. So
>>>> we enable 63 rx queues. But we found the rate of interrupt growth
>>>> on cpu 0~15 is faster than other cpus(almost twice).
>>> ...
>>>> I am confused that why ixgbe NIC can dispatch the packets
>>>> to the rx queues that not specified in RSS configuration.
>>>
>>> Hypothesis: it isn't doing so, RX is only happening on cpus (and
>>> queues) 0-15, but the other CPUs are still sending traffic and
>>> thus getting TX completion interrupts from their TX queues.
>>> `ethtool -S` output has per-queue traffic stats which should
>>> confirm this.
>>>
>>
>> I use ethtool -S to check the rx_queus stats and here is the result.
>>
>> According to the below stats, all cpus have new packets received.
>
> + Alex and Piotr
>
> What's your ntuple filter setting? If it's off, I suspect it may be the Flow Director ATR (Application Targeting Routing) feature which will utilize all queues. I believe if you turn on ntuple filters this will turn that feature off.
Yes, our ntuple filter setting is off. After turning on the ntuple filters, I compare the delta of recieved packets,
only 0~15 rx rings are non-zero, other rx rings are zero.
If we want to spread the packets across 0~62, how can we tune the NIC setting?
we have enabled 63 rx queues, irq_affinity and rx-flow-hash, but the 0~15 cpu
received more packets than others.
Thanks!
>
> Thanks,
> Tony
>
>>
>> cpu t1(bytes) t2(bytes) delta(bytes)
>>
>> 0 154155550267550 154156433828875 883561325
>> 1 148748566285840 148749509346247 943060407
>> 2 148874911191685 148875798038140 886846455
>> 3 152483460327704 152484251468998 791141294
>> 4 147790981836915 147791775847804 794010889
>> 5 146047892285722 146048778285682 885999960
>> 6 142880516825921 142881213804363 696978442
>> 7 152016735168735 152017707542774 972374039
>> 8 146019936404393 146020739070311 802665918
>> 9 147448522715540 147449258018186 735302646
>> 10 145865736299432 145866601503106 865203674
>> 11 149548527982122 149549289026453 761044331
>> 12 146848384328236 146849303547769 919219533
>> 13 152942139118542 152942769029253 629910711
>> 14 150884661854828 150885556866976 895012148
>> 15 149222733506734 149223510491115 776984381
>> 16 34150226069524 34150375855113 149785589
>> 17 34115700500819 34115914271025 213770206
>> 18 33906215129998 33906448044501 232914503
>> 19 33983812095357 33983986258546 174163189
>> 20 34156349675011 34156565159083 215484072
>> 21 33574293379024 33574490695725 197316701
>> 22 33438129453422 33438297911151 168457729
>> 23 32967454521585 32967612494711 157973126
>> 24 33507443427266 33507604828468 161401202
>> 25 33413275870121 33413433901940 158031819
>> 26 33852322542796 33852527061150 204518354
>> 27 33131162685385 33131330621474 167936089
>> 28 33407661780251 33407823112381 161332130
>> 29 34256799173845 34256944837757 145663912
>> 30 33814458585183 33814623673528 165088345
>> 31 33848638714862 33848775218038 136503176
>> 32 18683932398308 18684069540891 137142583
>> 33 19454524281229 19454647908293 123627064
>> 34 19717744365436 19717900618222 156252786
>> 35 20295086765202 20295245869666 159104464
>> 36 20501853066588 20502000738936 147672348
>> 37 20954631043374 20954797204375 166161001
>> 38 21102911073326 21103062510369 151437043
>> 39 21376404644179 21376515307288 110663109
>> 40 20935812784743 20935983891491 171106748
>> 41 20721278456831 20721435955715 157498884
>> 42 21268291801465 21268425244578 133443113
>> 43 21661413672829 21661629019091 215346262
>> 44 21696437732484 21696568800049 131067565
>> 45 21027869000890 21028020401214 151400324
>> 46 21707137252644 21707293761990 156509346
>> 47 20655623913790 20655740452889 116539099
>> 48 32692002128477 32692138244468 136115991
>> 49 33548445851486 33548569927672 124076186
>> 50 33197264968787 33197448645817 183677030
>> 51 33379544010500 33379746565576 202555076
>> 52 33503579011721 33503722596159 143584438
>> 53 33145734550468 33145892305819 157755351
>> 54 33422692741858 33422844156764 151414906
>> 55 32750945531107 32751131302251 185771144
>> 56 33404955373530 33405157766253 202392723
>> 57 33701185654471 33701313174725 127520254
>> 58 33014531699810 33014700058409 168358599
>> 59 32948906758429 32949151147605 244389176
>> 60 33470813725985 33470993164755 179438770
>> 61 33803771479735 33803971758441 200278706
>> 62 33509751180818 33509926649969 175469151
>>
>> Thanks!
>>
>>> (But Eric is right that if you _want_ RX to use every CPU you
>>> should just change the indirection table.)
>>
>
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-08 3:36 ` Haifeng Xu
@ 2025-01-08 21:06 ` Tony Nguyen
2025-01-09 3:26 ` Haifeng Xu
0 siblings, 1 reply; 15+ messages in thread
From: Tony Nguyen @ 2025-01-08 21:06 UTC (permalink / raw)
To: Haifeng Xu, Edward Cree, Eric Dumazet, Aleksandr Loktionov,
Kwapulinski, Piotr
Cc: Przemek Kitszel, David S. Miller, Jakub Kicinski, Paolo Abeni,
linux-kernel, netdev, intel-wired-lan
On 1/7/2025 7:36 PM, Haifeng Xu wrote:
>
>
> On 2025/1/8 01:16, Tony Nguyen wrote:
...
>>
>> What's your ntuple filter setting? If it's off, I suspect it may be the Flow Director ATR (Application Targeting Routing) feature which will utilize all queues. I believe if you turn on ntuple filters this will turn that feature off.
>
> Yes, our ntuple filter setting is off. After turning on the ntuple filters, I compare the delta of recieved packets,
> only 0~15 rx rings are non-zero, other rx rings are zero.
>
> If we want to spread the packets across 0~62, how can we tune the NIC setting?
> we have enabled 63 rx queues, irq_affinity and rx-flow-hash, but the 0~15 cpu
> received more packets than others.
As Jakub mentioned earlier, HW RSS is only supported on this device for
16 queues. ATR will steer bi-directional traffic to utilize additional
queues, however, once its exhausted it will fallback to RSS, which is
why CPUs 0-15 are receiving more traffic than the others. I'm not aware
of a way to evenly spread the traffic beyond the 16 HW supported RSS
queues for this device.
Thanks,
Tony
> Thanks!
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [Question] ixgbe：Mechanism of RSS
2025-01-08 21:06 ` Tony Nguyen
@ 2025-01-09 3:26 ` Haifeng Xu
0 siblings, 0 replies; 15+ messages in thread
From: Haifeng Xu @ 2025-01-09 3:26 UTC (permalink / raw)
To: Tony Nguyen, Edward Cree, Eric Dumazet, Aleksandr Loktionov,
Kwapulinski, Piotr
Cc: Przemek Kitszel, David S. Miller, Jakub Kicinski, Paolo Abeni,
linux-kernel, netdev, intel-wired-lan
On 2025/1/9 05:06, Tony Nguyen wrote:
>
>
> On 1/7/2025 7:36 PM, Haifeng Xu wrote:
>>
>>
>> On 2025/1/8 01:16, Tony Nguyen wrote:
>
> ...
>
>>>
>>> What's your ntuple filter setting? If it's off, I suspect it may be the Flow Director ATR (Application Targeting Routing) feature which will utilize all queues. I believe if you turn on ntuple filters this will turn that feature off.
>>
>> Yes, our ntuple filter setting is off. After turning on the ntuple filters, I compare the delta of recieved packets,
>> only 0~15 rx rings are non-zero, other rx rings are zero.
>>
>> If we want to spread the packets across 0~62, how can we tune the NIC setting?
>> we have enabled 63 rx queues, irq_affinity and rx-flow-hash, but the 0~15 cpu
>> received more packets than others.
>
> As Jakub mentioned earlier, HW RSS is only supported on this device for 16 queues. ATR will steer bi-directional traffic to utilize additional queues, however, once its exhausted it will fallback to RSS, which is why CPUs 0-15 are receiving more traffic than the others. I'm not aware of a way to evenly spread the traffic beyond the 16 HW supported RSS queues for this device.
Ok, thanks!
>
> Thanks,
> Tony
>
>> Thanks!
>
>
^ permalink raw reply [flat|nested] 15+ messages in thread
end of thread, other threads:[~2025-01-09 3:26 UTC | newest]
Thread overview: 15+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-01-02 3:53 [Question] ixgbe：Mechanism of RSS Haifeng Xu
2025-01-02 8:13 ` Eric Dumazet
2025-01-02 8:43 ` Haifeng Xu
2025-01-02 10:34 ` Eric Dumazet
2025-01-02 11:23 ` Haifeng Xu
2025-01-02 11:46 ` Eric Dumazet
2025-01-03 2:36 ` Haifeng Xu
2025-01-02 16:01 ` Edward Cree
2025-01-02 16:39 ` Jakub Kicinski
2025-01-03 2:37 ` Haifeng Xu
2025-01-03 3:05 ` Haifeng Xu
2025-01-07 17:16 ` Tony Nguyen
2025-01-08 3:36 ` Haifeng Xu
2025-01-08 21:06 ` Tony Nguyen
2025-01-09 3:26 ` Haifeng Xu
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).