Re: [dpdk-users] scheduler issue

[Alex Kiselev <alex@therouter.net>]

On 2020-12-07 12:32, Singh, Jasvinder wrote:
>> -----Original Message-----
>> From: Alex Kiselev <alex@therouter.net>
>> Sent: Monday, December 7, 2020 10:46 AM
>> To: Singh, Jasvinder <jasvinder.singh@intel.com>
>> Cc: users@dpdk.org; Dumitrescu, Cristian 
>> <cristian.dumitrescu@intel.com>;
>> Dharmappa, Savinay <savinay.dharmappa@intel.com>
>> Subject: Re: [dpdk-users] scheduler issue
>> 
>> On 2020-12-07 11:00, Singh, Jasvinder wrote:
>> >> -----Original Message-----
>> >> From: users <users-bounces@dpdk.org> On Behalf Of Alex Kiselev
>> >> Sent: Friday, November 27, 2020 12:12 PM
>> >> To: users@dpdk.org
>> >> Cc: Dumitrescu, Cristian <cristian.dumitrescu@intel.com>
>> >> Subject: Re: [dpdk-users] scheduler issue
>> >>
>> >> On 2020-11-25 16:04, Alex Kiselev wrote:
>> >> > On 2020-11-24 16:34, Alex Kiselev wrote:
>> >> >> Hello,
>> >> >>
>> >> >> I am facing a problem with the scheduler library DPDK 18.11.10
>> >> >> with default scheduler settings (RED is off).
>> >> >> It seems like some of the pipes (last time it was 4 out of 600
>> >> >> pipes) start incorrectly dropping most of the traffic after a
>> >> >> couple of days of successful work.
>> >> >>
>> >> >> So far I've checked that there are no mbuf leaks or any other
>> >> >> errors in my code and I am sure that traffic enters problematic pipes.
>> >> >> Also switching a traffic in the runtime to pipes of another port
>> >> >> restores the traffic flow.
>> >> >>
>> >> >> Ho do I approach debugging this issue?
>> >> >>
>> >> >> I've added using rte_sched_queue_read_stats(), but it doesn't give
>> >> >> me counters that accumulate values (packet drops for example), it
>> >> >> gives me some kind of current values and after a couple of seconds
>> >> >> those values are reset to zero, so I can say nothing based on that API.
>> >> >>
>> >> >> I would appreciate any ideas and help.
>> >> >> Thanks.
>> >> >
>> >> > Problematic pipes had very low bandwidth limit (1 Mbit/s) and also
>> >> > there is an oversubscription configuration event at subport 0 of
>> >> > port
>> >> > 13 to which those pipes belongs and
>> >> CONFIG_RTE_SCHED_SUBPORT_TC_OV is
>> >> > disabled.
>> >> >
>> >> > Could a congestion at that subport be the reason of the problem?
>> >> >
>> >> > How much overhead and performance degradation will add enabling
>> >> > CONFIG_RTE_SCHED_SUBPORT_TC_OV feature?
>> >> >
>> >> > Configuration:
>> >> >
>> >> >   #
>> >> >   # QoS Scheduler Profiles
>> >> >   #
>> >> >   hqos add profile  1 rate    8 K size 1000000 tc period 40
>> >> >   hqos add profile  2 rate  400 K size 1000000 tc period 40
>> >> >   hqos add profile  3 rate  600 K size 1000000 tc period 40
>> >> >   hqos add profile  4 rate  800 K size 1000000 tc period 40
>> >> >   hqos add profile  5 rate    1 M size 1000000 tc period 40
>> >> >   hqos add profile  6 rate 1500 K size 1000000 tc period 40
>> >> >   hqos add profile  7 rate    2 M size 1000000 tc period 40
>> >> >   hqos add profile  8 rate    3 M size 1000000 tc period 40
>> >> >   hqos add profile  9 rate    4 M size 1000000 tc period 40
>> >> >   hqos add profile 10 rate    5 M size 1000000 tc period 40
>> >> >   hqos add profile 11 rate    6 M size 1000000 tc period 40
>> >> >   hqos add profile 12 rate    8 M size 1000000 tc period 40
>> >> >   hqos add profile 13 rate   10 M size 1000000 tc period 40
>> >> >   hqos add profile 14 rate   12 M size 1000000 tc period 40
>> >> >   hqos add profile 15 rate   15 M size 1000000 tc period 40
>> >> >   hqos add profile 16 rate   16 M size 1000000 tc period 40
>> >> >   hqos add profile 17 rate   20 M size 1000000 tc period 40
>> >> >   hqos add profile 18 rate   30 M size 1000000 tc period 40
>> >> >   hqos add profile 19 rate   32 M size 1000000 tc period 40
>> >> >   hqos add profile 20 rate   40 M size 1000000 tc period 40
>> >> >   hqos add profile 21 rate   50 M size 1000000 tc period 40
>> >> >   hqos add profile 22 rate   60 M size 1000000 tc period 40
>> >> >   hqos add profile 23 rate  100 M size 1000000 tc period 40
>> >> >   hqos add profile 24 rate 25 M size 1000000 tc period 40
>> >> >   hqos add profile 25 rate 50 M size 1000000 tc period 40
>> >> >
>> >> >   #
>> >> >   # Port 13
>> >> >   #
>> >> >   hqos add port 13 rate 40 G mtu 1522 frame overhead 24 queue sizes
>> >> > 64
>> >> > 64 64 64
>> >> >   hqos add port 13 subport 0 rate 1500 M size 1000000 tc period 10
>> >> >   hqos add port 13 subport 0 pipes 3000 profile 2
>> >> >   hqos add port 13 subport 0 pipes 3000 profile 5
>> >> >   hqos add port 13 subport 0 pipes 3000 profile 6
>> >> >   hqos add port 13 subport 0 pipes 3000 profile 7
>> >> >   hqos add port 13 subport 0 pipes 3000 profile 9
>> >> >   hqos add port 13 subport 0 pipes 3000 profile 11
>> >> >   hqos set port 13 lcore 5
>> >>
>> >> I've enabled TC_OV feature and redirected most of the traffic to TC3.
>> >> But the issue still exists.
>> >>
>> >> Below is queue statistics of one of problematic pipes.
>> >> Almost all of the traffic entering the pipe is dropped.
>> >>
>> >> And the pipe is also configured with the 1Mbit/s profile.
>> >> So, the issue is only with very low bandwidth pipe profiles.
>> >>
>> >> And this time there was no congestion on the subport.
>> >>
>> >>
>> >> Egress qdisc
>> >> dir 0
>> >>    rate 1M
>> >>    port 6, subport 0, pipe_id 138, profile_id 5
>> >>    tc 0, queue 0: bytes 752, bytes dropped 0, pkts 8, pkts dropped 0
>> >>    tc 0, queue 1: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 0, queue 2: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 0, queue 3: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 1, queue 0: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 1, queue 1: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 1, queue 2: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 1, queue 3: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 2, queue 0: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 2, queue 1: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 2, queue 2: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 2, queue 3: bytes 0, bytes dropped 0, pkts 0, pkts dropped 0
>> >>    tc 3, queue 0: bytes 56669, bytes dropped 360242, pkts 150, pkts
>> >> dropped
>> >> 3749
>> >>    tc 3, queue 1: bytes 63005, bytes dropped 648782, pkts 150, pkts
>> >> dropped
>> >> 3164
>> >>    tc 3, queue 2: bytes 9984, bytes dropped 49704, pkts 128, pkts
>> >> dropped
>> >> 636
>> >>    tc 3, queue 3: bytes 15436, bytes dropped 107198, pkts 130, pkts
>> >> dropped
>> >> 354
>> >
>> >
>> > Hi Alex,
>> >
>> > Can you try newer version of the library, say dpdk 20.11?
>> 
>> Right now no, since switching to another DPDK will take a lot of time 
>> because
>> I am using a lot of custom patches.
>> 
>> I've tried to simply copy the entire rte_sched lib from DPDK 19 to 
>> DPDK 18.
>> And I was able to successful back port and resolve all dependency 
>> issues, but
>> it also will take some time to test this approach.
>> 
>> 
>> > Are you
>> > using dpdk qos sample app or your own app?
>> 
>> My own app.
>> 
>> >> What are the packets size?
>> 
>> Application is used as BRAS/BNG server, so it's used to provide 
>> internet
>> access to residential customers. Therefore packet sizes are typical to 
>> the
>> internet and vary from 64 to 1500 bytes. Most of the packets are 
>> around
>> 1000 bytes.
>> 
>> >
>> > Couple of other things for clarification- 1. At what rate you are
>> > injecting the traffic to low bandwidth pipes?
>> 
>> Well, the rate vary also, there could be congestion on some pipes at 
>> some
>> date time.
>> 
>> But the problem is that once the problem occurs at a pipe or at some 
>> queues
>> inside the pipe, the pipe stops transmitting even when incoming 
>> traffic rate is
>> much lower than the pipe's rate.
>> 
>> > 2. How is traffic distributed among pipes and their traffic class?
>> 
>> I am using IPv4 TOS field to choose the TC and there is a tos2tc map.
>> Most of my traffic has 0 tos value which is mapped to TC3 inside my 
>> app.
>> 
>> Recently I've switched to a tos2map which maps all traffic to TC3 to 
>> see if it
>> solves the problem.
>> 
>> Packet distribution to queues is done using the formula (ipv4.src +
>> ipv4.dst) & 3
>> 
>> > 3. Can you try putting your own counters on those pipes queues which
>> > periodically show the #packets in the queues to understand the
>> > dynamics?
>> 
>> I will try.
>> 
>> P.S.
>> 
>> Recently I've got another problem with scheduler.
>> 
>> After enabling the TC_OV feature one of the ports stops transmitting.
>> All port's pipes were affected.
>> Port had only one support, and there were only pipes with 1 Mbit/s 
>> profile.
>> The problem was solved by adding a 10Mit/s profile to that port. Only 
>> after
>> that port's pipes started to transmit.
>> I guess it has something to do with calculating tc_ov_wm as it depends 
>> on the
>> maximum pipe rate.
>> 
>> I am gonna make a test lab and a test build to reproduce this.

I've made some tests and was able to reproduce the port configuration 
issue
using a test build of my app.

Tests showed that TC_OV feature works not correctly in DPDK 18.11, but 
there are workarounds.

I still can't reproduce my main problem which is random pipes stop 
transmitting.

Here are details:

All tests use the same test traffic generator that produce
10 traffic flows entering 10 different pipes of port 1 subport 0.
Only queue 0 of each pipe is used.
TX rate is 800 kbit/s. packet size is 800 byte.
Pipes rate are 1 Mbit/s. Subport 0 rate is 500 Mbit/s.


###
### test 1
###

Traffic generator is configured to use TC3.

Configuration:

   hqos add profile 27 rate 1 M size 1000000 tc period 40
   hqos add profile 23 rate  100 M size 1000000 tc period 40

   # qos test port
   hqos add port 1 rate 10 G mtu 1522 frame overhead 24 queue sizes 64 64 
64 64
   hqos add port 1 subport 0 rate 500 M size 1000000 tc period 10
   hqos add port 1 subport 0 pipes 2000 profile 27
   hqos set port 1 lcore 3

Results:

h5 ~ # rcli sh qos rcv
rcv 0: rx rate 641280, nb pkts 501, ind 1
rcv 1: rx rate 641280, nb pkts 501, ind 1
rcv 2: rx rate 641280, nb pkts 501, ind 1
rcv 3: rx rate 641280, nb pkts 501, ind 1
rcv 4: rx rate 641280, nb pkts 501, ind 1
rcv 5: rx rate 641280, nb pkts 501, ind 1
rcv 6: rx rate 641280, nb pkts 501, ind 1
rcv 7: rx rate 641280, nb pkts 501, ind 1
rcv 8: rx rate 641280, nb pkts 501, ind 1
rcv 9: rx rate 641280, nb pkts 501, ind 1

! BUG
! RX rate is lower then expected 800000 bit/s despite
that there is no congestion neither at subport
nor at pipes levels.

###
### test 2
###

Traffic generator is configured to use TC3.
!!! profile 23 has been added to the test port.

Configuration:

   hqos add profile 27 rate 1 M size 1000000 tc period 40
   hqos add profile 23 rate  100 M size 1000000 tc period 40

   # qos test port
   hqos add port 1 rate 10 G mtu 1522 frame overhead 24 queue sizes 64 64 
64 64
   hqos add port 1 subport 0 rate 500 M size 1000000 tc period 10
   hqos add port 1 subport 0 pipes 2000 profile 27
   hqos add port 1 subport 0 pipes 200 profile 23
   hqos set port 1 lcore 3

Results:

h5 ~ # rcli sh qos rcv
rcv 0: rx rate 798720, nb pkts 624, ind 1
rcv 1: rx rate 798720, nb pkts 624, ind 1
rcv 2: rx rate 798720, nb pkts 624, ind 1
rcv 3: rx rate 798720, nb pkts 624, ind 1
rcv 4: rx rate 798720, nb pkts 624, ind 1
rcv 5: rx rate 798720, nb pkts 624, ind 1
rcv 6: rx rate 798720, nb pkts 624, ind 1
rcv 7: rx rate 798720, nb pkts 624, ind 1
rcv 8: rx rate 798720, nb pkts 624, ind 1
rcv 9: rx rate 798720, nb pkts 624, ind 1

OK.
Receiving traffic is rate is equal to expected values.
So, just adding a pipes which are not being used solves the problem.

###
### test 3
###

!!! traffic generator uses TC 0, so tc_ov is not being used in this 
test.
profile 23 is not used.

Configuration without profile 23.

   hqos add profile 27 rate 1 M size 1000000 tc period 40
   hqos add profile 23 rate  100 M size 1000000 tc period 40

   # qos test port
   hqos add port 1 rate 10 G mtu 1522 frame overhead 24 queue sizes 64 64 
64 64
   hqos add port 1 subport 0 rate 500 M size 1000000 tc period 10
   hqos add port 1 subport 0 pipes 2000 profile 27
   hqos set port 1 lcore 3

Restuls:

h5 ~ # rcli sh qos rcv
rcv 0: rx rate 798720, nb pkts 624, ind 0
rcv 1: rx rate 798720, nb pkts 624, ind 0
rcv 2: rx rate 798720, nb pkts 624, ind 0
rcv 3: rx rate 798720, nb pkts 624, ind 0
rcv 4: rx rate 798720, nb pkts 624, ind 0
rcv 5: rx rate 798720, nb pkts 624, ind 0
rcv 6: rx rate 798720, nb pkts 624, ind 0
rcv 7: rx rate 798720, nb pkts 624, ind 0
rcv 8: rx rate 798720, nb pkts 624, ind 0
rcv 9: rx rate 798720, nb pkts 624, ind 0

OK.
Receiving traffic is rate is equal to expected values.

###
### test 4
###

Traffic generator is configured to use TC3.
no profile 23.
!! subport tc period has been changed from 10 to 5.

Configuration:

   hqos add profile 27 rate 1 M size 1000000 tc period 40
   hqos add profile 23 rate  100 M size 1000000 tc period 40

   # qos test port
   hqos add port 1 rate 10 G mtu 1522 frame overhead 24 queue sizes 64 64 
64 64
   hqos add port 1 subport 0 rate 500 M size 1000000 tc period 5
   hqos add port 1 subport 0 pipes 2000 profile 27
   hqos set port 1 lcore 3

Restuls:

rcv 0: rx rate 0, nb pkts 0, ind 1
rcv 1: rx rate 0, nb pkts 0, ind 1
rcv 2: rx rate 0, nb pkts 0, ind 1
rcv 3: rx rate 0, nb pkts 0, ind 1
rcv 4: rx rate 0, nb pkts 0, ind 1
rcv 5: rx rate 0, nb pkts 0, ind 1
rcv 6: rx rate 0, nb pkts 0, ind 1
rcv 7: rx rate 0, nb pkts 0, ind 1
rcv 8: rx rate 0, nb pkts 0, ind 1
rcv 9: rx rate 0, nb pkts 0, ind 1

! zero traffic

###
### test 5
###

Traffic generator is configured to use TC3.
profile 23 is enabled.
subport tc period has been changed from 10 to 5.

Configuration:

   hqos add profile 27 rate 1 M size 1000000 tc period 40
   hqos add profile 23 rate  100 M size 1000000 tc period 40

   # qos test port
   hqos add port 1 rate 10 G mtu 1522 frame overhead 24 queue sizes 64 64 
64 64
   hqos add port 1 subport 0 rate 500 M size 1000000 tc period 5
   hqos add port 1 subport 0 pipes 2000 profile 27
   hqos add port 1 subport 0 pipes 200 profile 23
   hqos set port 1 lcore 3

Restuls:

h5 ~ # rcli sh qos rcv
rcv 0: rx rate 800000, nb pkts 625, ind 1
rcv 1: rx rate 800000, nb pkts 625, ind 1
rcv 2: rx rate 800000, nb pkts 625, ind 1
rcv 3: rx rate 800000, nb pkts 625, ind 1
rcv 4: rx rate 800000, nb pkts 625, ind 1
rcv 5: rx rate 800000, nb pkts 625, ind 1
rcv 6: rx rate 800000, nb pkts 625, ind 1
rcv 7: rx rate 800000, nb pkts 625, ind 1
rcv 8: rx rate 800000, nb pkts 625, ind 1
rcv 9: rx rate 800000, nb pkts 625, ind 1

OK


> 
> 
> Does this problem exist when you disable oversubscription mode? Worth
> looking at grinder_tc_ov_credits_update() and grinder_credits_update()
> functions where tc_ov_wm is altered.
> 
> 
>> >
>> > Thanks,
>> > Jasvinder