From: sreenaath vasudevan [mailto:[email protected]]
Sent: Thursday, February 18, 2016 8:01 PM
To: Dumitrescu, Cristian <cristian.dumitrescu at intel.com>
Cc: dev at dpdk.org
Subject: Re: [dpdk-dev] DPDK-QoS - link sharing across classes
Hi Cristian
Thanks for detailed response.
Your solution works so long as I have four queues in my current implementation.
[Cristian] Yes, I was relying on you saying you can actually group Q0 and Q1
together, as well as Q2 and Q3, Q4 and Q5, Q6 and Q7.
Following are the two issues I have now
1. I have 8 queues in the current implementation. This means I need to map the
existing 8 queues to two sets of 4 queues across two different classes (C0 and
C1) in DPDK-QOs right? The problem with that approach is that queue weights are
not relative across classes. Is there a way to work around this?
[Cristian] You can, of course, change to code to implement 8 queues per pipe
traffic class, but this is not going to be a trivial exercise. With the current
implementation unmodified, you could simply map 4 queues to e.g. TC0 and 4
queues to e.g. TC1, provided that it makes sense to prioritize the 4 queues of
TC0 as higher priority than those 4 queues mapped to TC1, so basically you are
OK with having strict priority (up to an upper limit) between the 2 sets of 4
queues.
2. B/w redistribution -
a) The moment I map the current implementation's 8 queues across two
different classes (say C0 and C1), would remaining b/w be distributed across
the two classes C0 and C1? Is it true that in the current DPDK-QoS
implementation, unused b/w gets distributed only to the last class C3? Would
not un-used b/w from C0 come to C1?
[Cristian] With 4 queues mapped to TC0 and 4 queues mapped to TC1, you can set
TC0 rate to X% of pipe rate and TC1 rate to 100% of pipe rate. This is the idea
behind strict priority: the sum of TC rates is usually bigger than 100% of pipe
rate, but this is fine, as due to strict priority the lowest priority TC used
(which in this case is TC1, not TC3, as you are not using TC2 and TC3 at all)
only gets 100% of the pipe rate when no traffic from higher priority TCs
exists. In this case, the extremes are: (1) TC0 demand is high, so TC0 uses
everything up to X%, so TC1 can use a max of (100 ? X)% and (2) TC0 demand is
zero, in which case TC1 is free to use up to 100% of pipe rate. So the answers
to your 3 questions are: yes/no/yes.
b) In current DPDK QoS implemention, if C1 has un-used b/w would that get
used by C0? Or is it only "lower priority class (C3, more specifically) uses
the un-used b/w from higher priority classes (C0,C1,C2) ?
[Cristian] Due to strict priority, it does not make sense to think about higher
priority classes using whatever is not used by low priority classes: If high
priority classes have traffic to send, they will always be picked in the
detriment of lower ones; when the high priority classes hit their upper limit
rate, then they are not allowed to send more, otherwise the upper limit makes
no sense. So the bandwidth reuse concept makes sense only for low prio TCs to
reuse whatever is not used from high prio TCs. Usually, this is configured by
fully provisioning TC0 .. TC2 and setting TC3 to 100% of pipe rate, so TC3 can
use its rate plus whatever gets unused by TC0 .. TC2, so TC3 rate is between:
100% - (rate TC0 + rate TC1 + rate TC2) and 100% of pipe rate. As explained
above, this can be applicable to e.g TC1 as well when only TC0 and TC1 are
actually used. So the answers to your 2 questions are: no/no.
On Tue, Feb 16, 2016 at 2:46 PM, Dumitrescu, Cristian <cristian.dumitrescu at
intel.com<mailto:cristian.dumitrescu at intel.com>> wrote:
> -----Original Message-----
> From: dev [mailto:dev-bounces at dpdk.org<mailto:dev-bounces at dpdk.org>] On
> Behalf Of sreenaath
> vasudevan
> Sent: Tuesday, February 2, 2016 9:09 PM
> To: dev at dpdk.org<mailto:dev at dpdk.org>
> Subject: [dpdk-dev] DPDK-QoS - link sharing across classes
>
> Hi
> I currently have QoS implemented in hardware and I am thinking of using
> DPDK's QoS feature to move it to software.
> Currently in the hardware,Based on the 4 class per pipe and 4 queues per
> class limitation, I was thinking of using 4 classes in DPDK-QoS and spread
> out the 8 h/w queues across the 4 classes.
> Let me explain with an example. Currently, this is how the h/w queue is
> represented
> Q0 - 10% b/w
> Q1- 10% b/w
> Q2- 15% b/w
> Q3 - 15% b/w
> Q4 - 15% b/w
> Q5 - 15% b/w
> Q6 - 10% b/w
> Q7 - 10% b/w
>
> Translating the above config to DPDK-QoS, based on my application need, Q0
> and Q1 can be logically grouped under class0 with upper b/w = 20%; Q2, Q3,
> Q4, Q5 can be logically grouped under class2 with upper b/w = 60%; Q6 and
> Q7 can be logically grouped under class 3 with super b/w = 20%.
>
> However, in the h/w, link sharing is available across all the 8 queues.
> DPDK materials say link sharing "typically" is enabled for last class, in
> this case class2. However, I also want class 1 or class 0 to use the
> remaining bandwidth when class2 does not have any traffic and so on. Can
> this be done in DPDK ? Do we have a concept of min and max b/w guarantee
> at
> the class level in DPDK-QoS ?
Your requirements seem to be:
- minimal rate for each class (with rates fully provisioned, i.e. sum of rate
not exceeding 100%)
- avoid b/w waste by redistributing unused b/w to the traffic classes that
currently have traffic according to their relative weights
In our implementation, traffic classes are scheduled in strict priority (with
upper limit defined, to avoid starvation), so TC0 traffic is always prioritized
for the current pipe over TC1 .. TC3 traffic (as long as current pipe has TC0
traffic and pipe TC0 rate is not reached). Therefore, the traffic class
hierarchy level is not going to help you here.
On the other hand, if you map all your queues/traffic classes to the queues of
one of our pipe traffic classes (it does not matter which one of TC0 .. TC3 you
pick, as long as you pick just one), your requirements become possible, as we
have 4 queues per each pipe traffic class, and they are scheduled using
weighted fair queuing (byte-level weighted round robin). Simply map your class0
to our pipe TC0 queue 0 (weight of 20%), your class2 to our pipe TC0 queue 1
(weight of 60%) and your class3 to our pipe TC0 queue 2 (weight of 20%), with
our pipe TC0 queue 3 unused, which should work exactly what you need. Makes
sense?
Regards,
Cristian
>
> Thanks !
>
> --
> regards
> sreenaath
--
regards
sreenaath
