[dpdk-users] IP Pipeline QoS

[Dumitrescu, Cristian]

Hi BL,

My answers inline below:

> -----Original Message-----
> From: longtb5 at viettel.com.vn [mailto:longtb5 at viettel.com.vn]
> Sent: Saturday, September 23, 2017 9:00 AM
> To: users at dpdk.org
> Cc: Dumitrescu, Cristian <cristian.dumitrescu at intel.com>
> Subject: IP Pipeline QoS
> 
> 
> Hi,
> I am trying to build a QoS/Traffic management application using packet
> framework. The initial goal is to be able to  configure traffic flow for upto 1000
> users, *individually*, through the front end cmdline.

Makes sense, you can map each subscriber/user to its own pipe (L3 node in the hierarchy), which basically results in 16x queues per subscriber split into 4x traffic classes.

> Atm I'm looking at ip_pipeline's edge_router_downstream sample and the
> qos_sched app for starting point.

Yes, these are good starting points.

> I have a few questions:
> 
> 1. The traffic management pipeline in edge_router_downstream.cfg is
> configured as followed:
> 
> [PIPELINE2]
> type = PASS-THROUGH
> pktq_in = SWQ0 SWQ1 SWQ2 SWQ3 TM0 TM1 TM2 TM3
> pktq_out = TM0 TM1 TM2 TM3 SWQ4 SWQ5 SWQ6 SWQ7
> 
> I'm not exactly sure how this works. My thinking is that since this is a passthru
> table with no action, the output of SWQ0 gets connected
> to the input of TM0 and the output of TM0 gets connected to input of SWQ4,
> effectively route SWQ0 to SWQ4 through TM0. Is that correct?

Yes, you got it right.

> 
> 2. If that's the case, why don't we do it this way:
> 
> [PIPELINE1]
> type = ROUTING
> pktq_in = RXQ0.0 RXQ1.0 RXQ2.0 RXQ3.0
> pktq_out = TM0 TM1 TM2 TM3 SINK0
> 
> [PIPELINE2]
> type = PASS-THROUGH
> pktq_in = TM0 TM1 TM2 TM3
> pktq_out = TM0 TM1 TM2 TM3
> 
> [PIPELINE3]
> type = PASS-THROUGH
> pktq_in = TM0 TM1 TM2 TM3
> pktq_out = TXQ0.0 TXQ1.0 TXQ2.0 TXQ3.0
> 
> In other words, why do we need SWQs in this case? (and in general what is
> the typical use of SWQs?)
> 

Great question!

First, I think what you are trying to suggest looks more like this one below, as we need to have a single producer and consumer for each TM, right?

[PIPELINE1]
type = ROUTING
pktq_in = RXQ0.0 RXQ1.0 RXQ2.0 RXQ3.0
pktq_out = TM0 TM1 TM2 TM3 SINK0

[PIPELINE2]
type = PASS-THROUGH
pktq_in = TM0 TM1 TM2 TM3
pktq_out = TXQ0.0 TXQ1.0 TXQ2.0 TXQ3.0

Second, this approach only works when both of these pipelines are on the same CPU (logical) core, as the TM port underlying rte_sched object has the restriction that enque() and dequeue() for the same port must be executed by the same thread. So eliminating the SWQs is actually dangerous, as you might later decide to push the two pipelines to different CPU cores (which can be quickly done through the ip_pipeline config file). So keeping the SWQs allow treating the TMs as internal objects to their pipeline, hence better encapsulation.

Third, what is the benefit of saving some SWQs? If pipelines are on different CPU cores, then the SWQs are a must due to thread safety. If pipelines are on same CPU core, then the SWQ producer and consumer are the same thread, so SWQ enqueue/dequeue overhead is very small (L1 cache read/write), so eliminating them does not provide any real performance benefit.

Makes sense?

> 3. I understand the fast/slow table copy mechanism for querying/updating
> _tables_ through the front end. How should I go about querying/updating
> pipe profile, which are parts of TM _ports_ if I'm not mistaken. For example,
> to get/set the rate of tc 0 of pipe profile 0.
> Put it another way, how can I configure tm_profile.cfg interactively through
> the CLI?
> Is it even possible to configure TMs on-the-fly like that?
> 

Yes, it is possible to do on-the-fly updates to TM configuration. This is done by re-invoking rte_sched_subport/pipe_config() functions after TM init has been completed.

Unfortunately we don't have the CLI commands for this yet in ip_pipeline application, so you would have to write them yourself (straightforward).

> Thanks,
> BL

Regards,
Cristian