On 1/7/19, 7:36 AM, "Mattias Rönnblom" <mattias.ronnb...@ericsson.com> wrote:
On 2018-12-21 20:12, Venky Venkatesh wrote:
>
>
> On 12/21/18, 10:59 AM, "Mattias Rönnblom" <mattias.ronnb...@ericsson.com>
wrote:
>
> On 2018-12-21 19:34, Venky Venkatesh wrote:
> >
> >
> > On 12/21/18, 10:24 AM, "Mattias Rönnblom"
<mattias.ronnb...@ericsson.com> wrote:
> >
> > On 2018-12-21 06:13, Venky Venkatesh wrote:
> > > Hi,
> > > We are considering using a multi-process mode of the DPDK
with the event generators and consumers being spread across multiple processes
(on different cores). We are also considering using the DSW eventdev. Is the
DSW designed for such a use case? If so, are there some restrictions and
something specific that need to be done to make it work correctly?
> > >
> >
> > The purpose of an event device is to do dynamic load
balancing across
> > multiple cores. Using the DPDK multiple-process support, with
its
> > requirement of having unique, non-overlapping, core masks
works against
> > or even defeats this purpose.
> >
> > [VV]: I don’t understand your last sentence. Suppose I am having
multiple packet processing processes (each with a single thread and polling a
disjoint set of queues) and each linked to DSW. Each process would invoke the
enqueue which will be handled by the DSW linked to that process. Will the DSWs
across these processes "collaborate" to get load balancing across the processes?
> >
>
> If the processes are to collaborate, and process packets in the same
> pipeline, they will need to share an event device (for example, a DSW
> instance).
>
> However, if you put each of your pipeline stages into a process with
a
> single worker thread, you will not leave any room for an event
device to
> load balance, since every eventdev queue will have only a single
> consumer linked to it.
>
> [VV]: Sorry for the ambiguous terminology used by me -- queue (above)
referred to port queues and not eventdev queues. Additionally, consider a very
simple pipeline -- just 1 stage followed by transmit. Thus each process is
pulling packets out of the port queue, enqueue into local DSW, dequeue from
local DSW and running this 1 stage pipeline and transmitting. The role of
eventdev in this world is to load balance across the processes -- that is what
I meant by DSWs collaborate (since they need to exchange load information and
do migration handshake). Hope that clarifies. Pls let me know if this will work.
I'm not familiar with the details of DPDK multiprocess support, but I
think this should work. Again, the DSW instance needs to be shared, and
can't be local to the process in case you want to use it to load balance
across different DPDK processes.
All of the huge page memory is shared, and that's the only memory a DSW
event device is using (except for execution stacks of course, which of
course doesn't have to be shared).
[VV]: I had a question on the eventdev initialization API in the above
multi-process setting. The following are the objects and API to init each of
them. For each of these can you confirm whether it needs to be called in the
PRIMARY process only or even the SECONDARY process must call these. The reason
for the question is that the shared memory must be safely initialized once.
Event device itself: rte_event_dev_configure, rte_event_dev_start
Ports: rte_event_port_setup
Queues: rte_event_queue_setup
Port-Queue-Links: rte_event_port_link
Thanks
-Venky
