Bill Fischofer(Bill-Fischofer-Linaro) replied on github web page:

example/generator/odp_generator.c
line 87
@@ -563,44 +572,68 @@ static int create_pktio(const char *dev, odp_pool_t pool,
                return -1;
        }
 
-       if (num_rx_queues > capa.max_input_queues)
-               num_rx_queues = capa.max_input_queues;
+       if (num_rx_queues) {
+               pktin_mode = ODP_PKTIO_OP_MT_UNSAFE;
+               if (num_rx_queues > capa.max_input_queues) {
+                       num_rx_queues = capa.max_input_queues;
+                       pktin_mode = ODP_PKTIO_OP_MT;
+                       EXAMPLE_DBG("Warning: Force RX multithread safe mode "
+                                   "(slower)on %s\n",  dev);
+               }
 
-       odp_pktin_queue_param_init(&pktin_param);
-       pktin_param.num_queues = num_rx_queues;
-       pktin_param.queue_param.sched.sync = ODP_SCHED_SYNC_ATOMIC;
+               odp_pktin_queue_param_init(&pktin_param);
+               pktin_param.num_queues = num_rx_queues;
+               pktin_param.op_mode = pktin_mode;
+               if (sched)
+                       pktin_param.queue_param.sched.sync =
+                               ODP_SCHED_SYNC_ATOMIC;


Comment:
Might be good to have options for controlling the queue sync type here as 
`ODP_SCHED_SYNC_PARALLEL` should result in highest throughput, and 
`ODP_SCHED_SYNC_ORDERED` would be useful in testing performance of scheduler 
implementations (in theory should be better than `ODP_SCHED_SYNC_ATOMIC`).

Something to explore in another PR

> muvarov wrote
> ok


>> muvarov wrote
>> and why odp_pktin_recv_tmo() and not odp_pktin_recv() ?


>>> muvarov wrote
>>> why not ODP_PKTIN_WAIT?


>>>> muvarov wrote
>>>> not all events are packets.


>>>>> muvarov wrote
>>>>> ```
>>>>>  * @return Next highest priority event
>>>>>  * @retval ODP_EVENT_INVALID on timeout and no events available
>>>>> ```


>>>>>> muvarov wrote
>>>>>> just separate rx function for scheduler and on thread start you just 
>>>>>> select scheduler or direct.


>>>>>>> bogdanPricope wrote
>>>>>>> This will complicate this already over-complicated code: we may need to 
>>>>>>> decide between ultimate performance and feature richness. 


>>>>>>>> bogdanPricope wrote
>>>>>>>> No -  we need to print csum errors first.
>>>>>>>> This part was significantly changed in api-next (csum checks use 
>>>>>>>> different/ new API) and it makes no sense to optimize it for the old 
>>>>>>>> (master) code. After integration in api-next, this part will be  
>>>>>>>> reworked  to use less parser flags (reduce parsing level).
>>>>>>>> For example, removing L4 parsing and locating interface is bringing an 
>>>>>>>> extra 1 mpps.


>>>>>>>>> bogdanPricope wrote
>>>>>>>>> '-r' may work.


>>>>>>>>>> Bill Fischofer(Bill-Fischofer-Linaro) wrote:
>>>>>>>>>> Having an option to use direct mode seems reasonable, but shouldn't 
>>>>>>>>>> we retain schedule mode (perhaps as a command line switch)? This 
>>>>>>>>>> would provide an easy means of testing scheduler efficiency as it is 
>>>>>>>>>> tuned. At least in some environments we'd like schedule mode to show 
>>>>>>>>>> better performance than direct.


>>>>>>>>>>> muvarov wrote
>>>>>>>>>>> that has to be the first check.


>>>>>>>>>>>> muvarov wrote
>>>>>>>>>>>> -r ?


https://github.com/Linaro/odp/pull/343#discussion_r158190837
updated_at 2017-12-21 03:50:22

Reply via email to