[RFC,v1,4/6] graph: enhance graph walk by cross-core dispatch
Checks
Commit Message
This patch enhance the task scheduler mechanism to enable dispatching
tasks to another worker cores. Currently, there is only a local work
queue for one graph to walk. We introduce a scheduler worker queue in
each worker core for dispatching tasks. It will perform the walk on
scheduler work queue first, then handle the local work queue.
Signed-off-by: Haiyue Wang <haiyue.wang@intel.com>
Signed-off-by: Cunming Liang <cunming.liang@intel.com>
Signed-off-by: Zhirun Yan <zhirun.yan@intel.com>
---
lib/graph/graph.c | 6 ++++++
lib/graph/rte_graph_worker.h | 11 +++++++++++
2 files changed, 17 insertions(+)
Comments
> This patch enhance the task scheduler mechanism to enable dispatching
> tasks to another worker cores. Currently, there is only a local work
> queue for one graph to walk. We introduce a scheduler worker queue in
> each worker core for dispatching tasks. It will perform the walk on
> scheduler work queue first, then handle the local work queue.
>
> Signed-off-by: Haiyue Wang <haiyue.wang@intel.com>
> Signed-off-by: Cunming Liang <cunming.liang@intel.com>
> Signed-off-by: Zhirun Yan <zhirun.yan@intel.com>
> ---
> lib/graph/graph.c | 6 ++++++
> lib/graph/rte_graph_worker.h | 11 +++++++++++
> 2 files changed, 17 insertions(+)
>
> diff --git a/lib/graph/graph.c b/lib/graph/graph.c
> index b4eb18175a..49ea2b3fbb 100644
> --- a/lib/graph/graph.c
> +++ b/lib/graph/graph.c
> @@ -368,6 +368,8 @@ rte_graph_destroy(rte_graph_t id)
> while (graph != NULL) {
> tmp = STAILQ_NEXT(graph, next);
> if (graph->id == id) {
> + /* Destroy the schedule work queue if has */
> + graph_sched_wq_destroy(graph);
> /* Call fini() of the all the nodes in the graph */
> graph_node_fini(graph);
> /* Destroy graph fast path memory */
> @@ -470,6 +472,10 @@ graph_clone(struct graph *parent_graph, const char
> *name,
> if (graph_node_init(graph))
> goto graph_mem_destroy;
>
> + /* Create the graph schedule work queue */
> + if (graph_sched_wq_create(graph, parent_graph))
> + goto graph_mem_destroy;
> +
> /* All good, Lets add the graph to the list */
> graph_id++;
> STAILQ_INSERT_TAIL(&graph_list, graph, next);
> diff --git a/lib/graph/rte_graph_worker.h b/lib/graph/rte_graph_worker.h
> index faf3f31ddc..e98697d880 100644
> --- a/lib/graph/rte_graph_worker.h
> +++ b/lib/graph/rte_graph_worker.h
> @@ -177,6 +177,7 @@ static inline void
> rte_graph_walk(struct rte_graph *graph)
> {
> const rte_graph_off_t *cir_start = graph->cir_start;
> + const unsigned int lcore_id = graph->lcore_id;
> const rte_node_t mask = graph->cir_mask;
> uint32_t head = graph->head;
> struct rte_node *node;
> @@ -184,6 +185,9 @@ rte_graph_walk(struct rte_graph *graph)
> uint16_t rc;
> void **objs;
>
> + if (graph->wq != NULL)
> + __rte_graph_sched_wq_process(graph);
> +
We should introduce a flags field in rte_graph_param which can
be used by the application to define whether a graph should support
multi-core dispatch.
Then we can make `__rte_graph_sched_wq_process` as node 0 during graph
creation so that it will be always called at the start of graph processing followed
by calling rest of the nodes.
This will remove unnecessary branches in fastpath.
> /*
> * Walk on the source node(s) ((cir_start - head) -> cir_start) and
> then
> * on the pending streams (cir_start -> (cir_start + mask) -> cir_start)
> @@ -205,6 +209,12 @@ rte_graph_walk(struct rte_graph *graph)
> objs = node->objs;
> rte_prefetch0(objs);
>
> + /* Schedule the node until all task/objs are done */
> + if (node->lcore_id != RTE_MAX_LCORE && (int32_t)head > 0
> &&
> + lcore_id != node->lcore_id && graph->rq != NULL &&
> + __rte_graph_sched_node_enqueue(node, graph->rq))
> + goto next;
> +
> if (rte_graph_has_stats_feature()) {
> start = rte_rdtsc();
> rc = node->process(graph, node, objs, node->idx);
> @@ -215,6 +225,7 @@ rte_graph_walk(struct rte_graph *graph)
> node->process(graph, node, objs, node->idx);
> }
> node->idx = 0;
> + next:
> head = likely((int32_t)head > 0) ? head & mask : head;
> }
> graph->tail = 0;
> --
> 2.25.1
> -----Original Message-----
> From: Pavan Nikhilesh Bhagavatula <pbhagavatula@marvell.com>
> Sent: Thursday, September 8, 2022 1:27 PM
> To: Yan, Zhirun <zhirun.yan@intel.com>; dev@dpdk.org; Jerin Jacob
> Kollanukkaran <jerinj@marvell.com>; Kiran Kumar Kokkilagadda
> <kirankumark@marvell.com>
> Cc: Liang, Cunming <cunming.liang@intel.com>; Wang, Haiyue
> <haiyue.wang@intel.com>
> Subject: RE: [EXT] [RFC, v1 4/6] graph: enhance graph walk by cross-core
> dispatch
>
> > This patch enhance the task scheduler mechanism to enable dispatching
> > tasks to another worker cores. Currently, there is only a local work
> > queue for one graph to walk. We introduce a scheduler worker queue in
> > each worker core for dispatching tasks. It will perform the walk on
> > scheduler work queue first, then handle the local work queue.
> >
> > Signed-off-by: Haiyue Wang <haiyue.wang@intel.com>
> > Signed-off-by: Cunming Liang <cunming.liang@intel.com>
> > Signed-off-by: Zhirun Yan <zhirun.yan@intel.com>
> > ---
> > lib/graph/graph.c | 6 ++++++
> > lib/graph/rte_graph_worker.h | 11 +++++++++++
> > 2 files changed, 17 insertions(+)
> >
> > diff --git a/lib/graph/graph.c b/lib/graph/graph.c index
> > b4eb18175a..49ea2b3fbb 100644
> > --- a/lib/graph/graph.c
> > +++ b/lib/graph/graph.c
> > @@ -368,6 +368,8 @@ rte_graph_destroy(rte_graph_t id)
> > while (graph != NULL) {
> > tmp = STAILQ_NEXT(graph, next);
> > if (graph->id == id) {
> > + /* Destroy the schedule work queue if has */
> > + graph_sched_wq_destroy(graph);
> > /* Call fini() of the all the nodes in the graph */
> > graph_node_fini(graph);
> > /* Destroy graph fast path memory */ @@ -470,6 +472,10
> @@
> > graph_clone(struct graph *parent_graph, const char *name,
> > if (graph_node_init(graph))
> > goto graph_mem_destroy;
> >
> > + /* Create the graph schedule work queue */
> > + if (graph_sched_wq_create(graph, parent_graph))
> > + goto graph_mem_destroy;
> > +
> > /* All good, Lets add the graph to the list */
> > graph_id++;
> > STAILQ_INSERT_TAIL(&graph_list, graph, next); diff --git
> > a/lib/graph/rte_graph_worker.h b/lib/graph/rte_graph_worker.h index
> > faf3f31ddc..e98697d880 100644
> > --- a/lib/graph/rte_graph_worker.h
> > +++ b/lib/graph/rte_graph_worker.h
> > @@ -177,6 +177,7 @@ static inline void rte_graph_walk(struct
> > rte_graph *graph) {
> > const rte_graph_off_t *cir_start = graph->cir_start;
> > + const unsigned int lcore_id = graph->lcore_id;
> > const rte_node_t mask = graph->cir_mask;
> > uint32_t head = graph->head;
> > struct rte_node *node;
> > @@ -184,6 +185,9 @@ rte_graph_walk(struct rte_graph *graph)
> > uint16_t rc;
> > void **objs;
> >
> > + if (graph->wq != NULL)
> > + __rte_graph_sched_wq_process(graph);
> > +
>
>
> We should introduce a flags field in rte_graph_param which can be used by
> the application to define whether a graph should support multi-core
> dispatch.
>
Yes, I will add a flags field in next version.
> Then we can make `__rte_graph_sched_wq_process` as node 0 during graph
> creation so that it will be always called at the start of graph processing
> followed by calling rest of the nodes.
> This will remove unnecessary branches in fastpath.
>
Thanks for your comments and sorry for my late reply.
Yes, we can make `__rte_graph_sched_wq_process` as node 0 with dispatch flags.
But I am not sure whether we need to register a new node here. It means we should change
the graph topo, and it will be an isolated node with no links.
> > /*
> > * Walk on the source node(s) ((cir_start - head) -> cir_start) and
> > then
> > * on the pending streams (cir_start -> (cir_start + mask) ->
> > cir_start) @@ -205,6 +209,12 @@ rte_graph_walk(struct rte_graph
> *graph)
> > objs = node->objs;
> > rte_prefetch0(objs);
> >
> > + /* Schedule the node until all task/objs are done */
> > + if (node->lcore_id != RTE_MAX_LCORE && (int32_t)head > 0
> > &&
> > + lcore_id != node->lcore_id && graph->rq != NULL &&
> > + __rte_graph_sched_node_enqueue(node, graph->rq))
> > + goto next;
> > +
> > if (rte_graph_has_stats_feature()) {
> > start = rte_rdtsc();
> > rc = node->process(graph, node, objs, node->idx); @@ -215,6
> +225,7
> > @@ rte_graph_walk(struct rte_graph *graph)
> > node->process(graph, node, objs, node->idx);
> > }
> > node->idx = 0;
> > + next:
> > head = likely((int32_t)head > 0) ? head & mask : head;
> > }
> > graph->tail = 0;
> > --
> > 2.25.1
@@ -368,6 +368,8 @@ rte_graph_destroy(rte_graph_t id)
while (graph != NULL) {
tmp = STAILQ_NEXT(graph, next);
if (graph->id == id) {
+ /* Destroy the schedule work queue if has */
+ graph_sched_wq_destroy(graph);
/* Call fini() of the all the nodes in the graph */
graph_node_fini(graph);
/* Destroy graph fast path memory */
@@ -470,6 +472,10 @@ graph_clone(struct graph *parent_graph, const char *name,
if (graph_node_init(graph))
goto graph_mem_destroy;
+ /* Create the graph schedule work queue */
+ if (graph_sched_wq_create(graph, parent_graph))
+ goto graph_mem_destroy;
+
/* All good, Lets add the graph to the list */
graph_id++;
STAILQ_INSERT_TAIL(&graph_list, graph, next);
@@ -177,6 +177,7 @@ static inline void
rte_graph_walk(struct rte_graph *graph)
{
const rte_graph_off_t *cir_start = graph->cir_start;
+ const unsigned int lcore_id = graph->lcore_id;
const rte_node_t mask = graph->cir_mask;
uint32_t head = graph->head;
struct rte_node *node;
@@ -184,6 +185,9 @@ rte_graph_walk(struct rte_graph *graph)
uint16_t rc;
void **objs;
+ if (graph->wq != NULL)
+ __rte_graph_sched_wq_process(graph);
+
/*
* Walk on the source node(s) ((cir_start - head) -> cir_start) and then
* on the pending streams (cir_start -> (cir_start + mask) -> cir_start)
@@ -205,6 +209,12 @@ rte_graph_walk(struct rte_graph *graph)
objs = node->objs;
rte_prefetch0(objs);
+ /* Schedule the node until all task/objs are done */
+ if (node->lcore_id != RTE_MAX_LCORE && (int32_t)head > 0 &&
+ lcore_id != node->lcore_id && graph->rq != NULL &&
+ __rte_graph_sched_node_enqueue(node, graph->rq))
+ goto next;
+
if (rte_graph_has_stats_feature()) {
start = rte_rdtsc();
rc = node->process(graph, node, objs, node->idx);
@@ -215,6 +225,7 @@ rte_graph_walk(struct rte_graph *graph)
node->process(graph, node, objs, node->idx);
}
node->idx = 0;
+ next:
head = likely((int32_t)head > 0) ? head & mask : head;
}
graph->tail = 0;