[dpdk-dev] [PATCH 2/2] event/sw: use dynamically-sized IQs

[Van Haaren, Harry]

> From: Eads, Gage
> Sent: Thursday, November 30, 2017 3:09 AM
> To: dev at dpdk.org
> Cc: jerin.jacob at caviumnetworks.com; Van Haaren, Harry
> <harry.van.haaren at intel.com>; Richardson, Bruce
> <bruce.richardson at intel.com>; hemant.agrawal at nxp.com; nipun.gupta at nxp.com;
> santosh.shukla at caviumnetworks.com; pbhagavatula at caviumnetworks.com
> Subject: [PATCH 2/2] event/sw: use dynamically-sized IQs
> 
> This commit introduces dynamically-sized IQs, by switching the underlying
> data structure from a fixed-size ring to a linked list of queue 'chunks.'
> This has a number of benefits:
> - Certain corner cases were observed in which all of a pipeline's flows
>   could be pinned to one port for extended periods, effectively turning a
>   multi-core pipeline into single-core one. This was caused by an event
>   producer having a larger new_event_threshold than the IQ depth, and
>   injecting large numbers of packets that are ultimately backpressured in a
>   worker's rx_ring, causing those packets' flows to be scheduled to that
>   port.
>   The dynamically sized IQ does not have this problem because each IQ can
>   grow large enough to store all the system's events, such that
>   backpressure will not reach the worker_ring.
> - Slight performance improvement (~1-2%) in high throughput scenarios,
>   tested with eventdev_pipeline_sw_pmd.
> 
> This implementation has a small increase in the queue storage memory
> footprint (~70KB). This commit also removes the iq_size xstat, which no
> longer applies to this implementation.
> 
> Signed-off-by: Gage Eads <gage.eads at intel.com>

Some review notes below - but nothing that needs changing.

Acked-by: Harry van Haaren <harry.van.haaren at intel.com>


> diff --git a/drivers/event/sw/sw_evdev.c b/drivers/event/sw/sw_evdev.c
<snip>

> -	for (i = 0; i < SW_IQS_MAX; i++) {
> -		snprintf(buf, sizeof(buf), "q_%u_iq_%d", idx, i);
> -		qid->iq[i] = iq_ring_create(buf, socket_id);
> -		if (!qid->iq[i]) {
> -			SW_LOG_DBG("ring create failed");
> -			goto cleanup;
> -		}
> -	}
> +	for (i = 0; i < SW_IQS_MAX; i++)
> +		iq_init(sw, &qid->iq[i]);

Review notes;
- It looks like error checking is being lost above, but in reality the iq_init() cannot fail in this new scheme, provided that the pool of chunks is provisioned correctly at configure() time (see provisioning below).

No action required here - just noting this error-checking change.


> +	/* Number of chunks sized for worst-case spread of events across IQs */
> +	num_chunks = ((SW_INFLIGHT_EVENTS_TOTAL/SW_EVS_PER_Q_CHUNK)+1) +
> +			sw->qid_count*SW_IQS_MAX*2;

To verify the logic here;

((TOTAL / CHUNK_SIZE) + 1) is the main body of chunks.
QID * IQ * 2 is there so each QID can hold a chunk, and a next pointer, for each priority of IQ.

add these two together, is the total number of chunks. Check :)