[PATCH v17] graph: add optional profiling stats
kirankumark at marvell.com
kirankumark at marvell.com
Mon Aug 24 13:15:51 CEST 2026
On Sun, Aug 23, 2026 at 03:01:27PM +0000, Morten Brørup wrote:
> Added graph node profiling stats, build time configurable by enabling
> RTE_GRAPH_PROFILE in rte_config.h.
>
> Signed-off-by: Morten Brørup <mb at smartsharesystems.com>
> ---
Acked-by: Kiran Kumar Kokkilagadda <kirankumark at marvell.com>
> v17:
> * Snapshot cycles as positive number, and subtract after processing.
> (Robin Jarry)
> v16:
> * Refactored range bucketing for clarity and performance. (AI)
> v15:
> * Used the term "ranges" instead of "intervals". (AI)
> * Described the last range as "%u+" instead of "[%u;[".
> Correct would be "[%u;+∞[", but prefer ASCII here.
> v14:
> * Changed some minor details suggested by AI.
> v13:
> * Fixed cache alignment; array should be aligned, not struct.
> * Changed some minor details suggested by AI.
> v12:
> * Multiply printed percent values by 100.0. (AI)
> * Added to release note that this feature is disabled by default. (AI)
> v11:
> * Fixed signedness comparison and copy-paste bug.
> v10:
> * Use intervals instead of point samples. (Robin Jarry)
> * Do not double-indent printf continuations.
> Follows existing coding style in the file, and
> avoids a long line warning from checkpatch.
> * Rebased.
> Moved release note to 26.11 file.
> v9:
> * Fixed comment still mentioning 32 objects.
> * Moved sample size array outside loop. (AI)
> * Added release note. (AI)
> v8:
> * Added static const array as local variable, instead of indexing directly
> into const array. (AI)
> This also eliminates the space required between "} [idx];" weirdness.
> * Added build time configurable RTE_GRAPH_PROFILE_BURST_SIZE to replace
> the hardcoded burst size of 32. (AI)
> v7:
> * Use RTE_DIM() in histogram for loop.
> * Added static_assert for histogram index values.
> * Minor details to please checkpatch.
> Although I disagree with requiring a space when indexing into
> a constant array "(const type []){values} [idx];",
> I have changed the code to comply.
> v6:
> * Consolidate the four histogram entries into one array. (Saeed Bishara)
> * Sample at 32 objs instead of a half burst. (Saeed Bishara)
> * Moved stats to different location in rte_node structure. (Jerin)
> * Minor details to please checkpatch.
> v5:
> * Added stats for a half burst and a full burst.
> v4:
> * Added documentation. (AI)
> * Added more comments. (AI)
> * Improved dump. (AI)
> * Debug shows both cycles/call and cycles/obj.
> v3:
> * Debug shows cycles/obj instead of cycles/call.
> * Fixed missing --in-reply-to.
> v2:
> * Fixed indentation.
> ---
> config/rte_config.h | 4 ++
> doc/guides/prog_guide/graph_lib.rst | 5 ++
> doc/guides/rel_notes/release_26_11.rst | 9 ++++
> lib/graph/graph_debug.c | 67 ++++++++++++++++++++++++++
> lib/graph/node.c | 2 +
> lib/graph/rte_graph_worker_common.h | 57 ++++++++++++++++++++--
> 6 files changed, 141 insertions(+), 3 deletions(-)
>
> diff --git a/config/rte_config.h b/config/rte_config.h
> index 0447cdf2ad..6beee01b09 100644
> --- a/config/rte_config.h
> +++ b/config/rte_config.h
> @@ -106,6 +106,10 @@
> /* rte_graph defines */
> #define RTE_GRAPH_BURST_SIZE 256
> #define RTE_LIBRTE_GRAPH_STATS 1
> +/* RTE_GRAPH_PROFILE is not set */
> +#define RTE_GRAPH_PROFILE_BURST_SMALL 8
> +#define RTE_GRAPH_PROFILE_BURST_MEDIUM 32
> +#define RTE_GRAPH_PROFILE_BURST_LARGE RTE_GRAPH_BURST_SIZE
>
> /****** driver defines ********/
>
> diff --git a/doc/guides/prog_guide/graph_lib.rst b/doc/guides/prog_guide/graph_lib.rst
> index 1c2cd1ef31..aa5faf4e82 100644
> --- a/doc/guides/prog_guide/graph_lib.rst
> +++ b/doc/guides/prog_guide/graph_lib.rst
> @@ -49,6 +49,11 @@ Performance tuning parameters
> RTE_GRAPH_BURST_SIZE config option.
> The testing shows, on x86 and arm64 servers, The sweet spot is 256 burst
> size. While on arm64 embedded SoCs, it is either 64 or 128.
> +- Enable the ``RTE_GRAPH_PROFILE`` config option for more profiling details
> + (disabled by default).
> + Set the ``RTE_GRAPH_PROFILE_BURST_SMALL``, ``RTE_GRAPH_PROFILE_BURST_MEDIUM``,
> + and ``RTE_GRAPH_PROFILE_BURST_LARGE`` config options to sample specific
> + burst size ranges.
> - Disable node statistics (using ``RTE_LIBRTE_GRAPH_STATS`` config option)
> if not needed.
>
> diff --git a/doc/guides/rel_notes/release_26_11.rst b/doc/guides/rel_notes/release_26_11.rst
> index c8cc86295d..2c822b6eee 100644
> --- a/doc/guides/rel_notes/release_26_11.rst
> +++ b/doc/guides/rel_notes/release_26_11.rst
> @@ -55,6 +55,15 @@ New Features
> Also, make sure to start the actual text at the margin.
> =======================================================
>
> +* **Added optional graph profiling statistics.**
> +
> + Added build time configurable graph node profiling statistics via
> + ``RTE_GRAPH_PROFILE`` in ``rte_config.h`` (disabled by default). When enabled,
> + tracks cycles spent processing various burst size ranges per node.
> + The edges of the ranges are build time configurable via
> + ``RTE_GRAPH_PROFILE_BURST_SMALL``, ``RTE_GRAPH_PROFILE_BURST_MEDIUM``, and
> + ``RTE_GRAPH_PROFILE_BURST_LARGE`` in ``rte_config.h``.
> +
>
> Removed Items
> -------------
> diff --git a/lib/graph/graph_debug.c b/lib/graph/graph_debug.c
> index e3b8cccdc1..a2aef9918f 100644
> --- a/lib/graph/graph_debug.c
> +++ b/lib/graph/graph_debug.c
> @@ -60,6 +60,7 @@ rte_graph_obj_dump(FILE *f, struct rte_graph *g, bool all)
> rte_graph_off_t off;
> struct rte_node *n;
> rte_edge_t i;
> + uint64_t all_total_cycles = 0;
>
> fprintf(f, "graph <%s> @ %p\n", g->name, g);
> fprintf(f, " id=%" PRIu32 "\n", g->id);
> @@ -72,6 +73,12 @@ rte_graph_obj_dump(FILE *f, struct rte_graph *g, bool all)
> fprintf(f, " nodes_start=0x%" PRIx32 "\n", g->nodes_start);
> fprintf(f, " cir_start=%p\n", g->cir_start);
>
> + rte_graph_foreach_node(count, off, g, n) {
> + if (!all && n->idx == 0)
> + continue;
> + all_total_cycles += n->total_cycles;
> + }
> +
> rte_graph_foreach_node(count, off, g, n) {
> if (!all && n->idx == 0)
> continue;
> @@ -93,6 +100,66 @@ rte_graph_obj_dump(FILE *f, struct rte_graph *g, bool all)
> n->dispatch.total_sched_fail);
> }
> fprintf(f, " total_calls=%" PRId64 "\n", n->total_calls);
> + if (rte_graph_has_stats_feature())
> + fprintf(f, " total_cycles=%" PRIu64 " (%.1f%% of all nodes), avg cycles/call=%.1f\n",
> + n->total_cycles,
> + all_total_cycles == 0 ? 0.0 :
> + (double)n->total_cycles / (double)all_total_cycles * 100.0,
> + n->total_calls == 0 ? 0.0 :
> + (double)n->total_cycles / (double)n->total_calls);
> +
> +#ifdef RTE_GRAPH_PROFILE
> + static const uint16_t sample_sizes[] = {
> + 0, 1, 2, RTE_GRAPH_PROFILE_BURST_SMALL,
> + RTE_GRAPH_PROFILE_BURST_MEDIUM, RTE_GRAPH_PROFILE_BURST_LARGE};
> + static_assert(RTE_DIM(sample_sizes) == 2U + RTE_DIM(n->usage_stats),
> + "usage_stats array size mismatch");
> + for (unsigned int idx = 0; idx < RTE_DIM(sample_sizes); idx++) {
> + uint64_t calls;
> + uint64_t cycles;
> + double objs_per_call;
> + if (idx < 2) {
> + calls = n->usage_stats_01[idx].calls;
> + cycles = n->usage_stats_01[idx].cycles;
> + objs_per_call = (double)idx;
> + fprintf(f, " for %u objs/call\n",
> + idx);
> + } else {
> + calls = n->usage_stats[idx - 2].calls;
> + cycles = n->usage_stats[idx - 2].cycles;
> + objs_per_call = calls == 0 ? 0.0 :
> + (double)n->usage_stats[idx - 2].objs /
> + (double)calls;
> + if (idx < RTE_DIM(sample_sizes) - 1)
> + fprintf(f, " for [%u;%u[ objs/call",
> + sample_sizes[idx], sample_sizes[idx + 1]);
> + else
> + fprintf(f, " for %u+ objs/call",
> + sample_sizes[idx]);
> + if (calls != 0)
> + fprintf(f, ", avg %.1f objs/call\n",
> + objs_per_call);
> + fprintf(f, "\n");
> + }
> + fprintf(f, " calls=%" PRIu64,
> + calls);
> + if (calls != 0) {
> + fprintf(f, " (%.1f%% of this node)\n",
> + n->total_calls == 0 ? 0.0 :
> + (double)calls / (double)n->total_calls * 100.0);
> + fprintf(f, " cycles=%" PRIu64 " (%.1f%% of this node), avg cycles/call=%.1f",
> + cycles,
> + n->total_cycles == 0 ? 0.0 :
> + (double)cycles / (double)n->total_cycles * 100.0,
> + (double)cycles / (double)calls);
> + if (objs_per_call != 0.0)
> + fprintf(f, ", avg cycles/obj=%.1f",
> + (double)cycles / (double)calls / objs_per_call);
> + }
> + fprintf(f, "\n");
> + }
> +#endif /* RTE_GRAPH_PROFILE */
> +
> for (i = 0; i < n->nb_edges; i++)
> fprintf(f, " edge[%d] <%s>\n", i,
> n->nodes[i]->name);
> diff --git a/lib/graph/node.c b/lib/graph/node.c
> index 1fce3e6632..19b38881ae 100644
> --- a/lib/graph/node.c
> +++ b/lib/graph/node.c
> @@ -110,10 +110,12 @@ __rte_node_register(const struct rte_node_register *reg)
> rte_edge_t i;
> size_t sz;
>
> +#ifndef RTE_GRAPH_PROFILE
> /* Limit Node specific metadata to one cacheline on 64B CL machine */
> RTE_BUILD_BUG_ON((offsetof(struct rte_node, nodes) -
> offsetof(struct rte_node, ctx)) !=
> RTE_CACHE_LINE_MIN_SIZE);
> +#endif
>
> graph_spinlock_lock();
>
> diff --git a/lib/graph/rte_graph_worker_common.h b/lib/graph/rte_graph_worker_common.h
> index 4ab53a533e..89c1f3dbf9 100644
> --- a/lib/graph/rte_graph_worker_common.h
> +++ b/lib/graph/rte_graph_worker_common.h
> @@ -121,6 +121,27 @@ struct __rte_cache_aligned rte_node {
> rte_graph_off_t xstat_off; /**< Offset to xstat counters. */
>
> /** Fast path area cache line 2. */
> +#ifdef RTE_GRAPH_PROFILE
> + /** Usage when this node processed 0 or 1 objects. */
> + alignas(RTE_CACHE_LINE_MIN_SIZE)
> + struct {
> + uint64_t calls; /**< Calls done. */
> + uint64_t cycles; /**< Cycles spent. */
> + } usage_stats_01[2];
> + /**
> + * Usage when this node processed N objects:
> + * - index 0 (micro): 2 <= N < RTE_GRAPH_PROFILE_BURST_SMALL
> + * - index 1 (small): RTE_GRAPH_PROFILE_BURST_SMALL <= N < RTE_GRAPH_PROFILE_BURST_MEDIUM
> + * - index 2 (medium): RTE_GRAPH_PROFILE_BURST_MEDIUM <= N < RTE_GRAPH_PROFILE_BURST_LARGE
> + * - index 3 (large): RTE_GRAPH_PROFILE_BURST_LARGE <= N
> + */
> + struct {
> + uint64_t calls; /**< Calls done. */
> + uint64_t cycles; /**< Cycles spent. */
> + uint64_t objs; /**< Objects processed. */
> + } usage_stats[4];
> + /** Fast path area cache line 4. */
> +#endif
> __extension__ struct __rte_cache_aligned {
> #define RTE_NODE_CTX_SZ 16
> union {
> @@ -148,8 +169,10 @@ struct __rte_cache_aligned rte_node {
> };
> };
>
> +#ifndef RTE_GRAPH_PROFILE
> static_assert(offsetof(struct rte_node, nodes) - offsetof(struct rte_node, ctx)
> == RTE_CACHE_LINE_MIN_SIZE, "rte_node fast path area must fit in 64 bytes");
> +#endif
>
> /**
> * @internal
> @@ -197,7 +220,7 @@ void __rte_node_stream_alloc_size(struct rte_graph *graph,
> static __rte_always_inline void
> __rte_node_process(struct rte_graph *graph, struct rte_node *node)
> {
> - uint64_t start;
> + uint64_t cycles;
> uint16_t rc;
> void **objs;
>
> @@ -206,11 +229,39 @@ __rte_node_process(struct rte_graph *graph, struct rte_node *node)
> rte_prefetch0(objs);
>
> if (rte_graph_has_stats_feature()) {
> - start = rte_rdtsc();
> + cycles = rte_rdtsc();
> rc = node->process(graph, node, objs, node->idx);
> - node->total_cycles += rte_rdtsc() - start;
> + cycles = rte_rdtsc() - cycles;
> + node->total_cycles += cycles;
> node->total_calls++;
> node->total_objs += rc;
> +#ifdef RTE_GRAPH_PROFILE
> + static_assert(RTE_GRAPH_PROFILE_BURST_SMALL >= 2,
> + "BURST_SMALL must be >= 2");
> + static_assert(RTE_GRAPH_PROFILE_BURST_SMALL < RTE_GRAPH_PROFILE_BURST_MEDIUM,
> + "BURST_SMALL must be < BURST_MEDIUM");
> + static_assert(RTE_GRAPH_PROFILE_BURST_MEDIUM < RTE_GRAPH_PROFILE_BURST_LARGE,
> + "BURST_MEDIUM must be < BURST_LARGE");
> + if (rc < 2) {
> + node->usage_stats_01[rc].calls++;
> + node->usage_stats_01[rc].cycles += cycles;
> + } else {
> + unsigned int index;
> +
> + if (rc >= RTE_GRAPH_PROFILE_BURST_LARGE)
> + index = 3;
> + else if (rc >= RTE_GRAPH_PROFILE_BURST_MEDIUM)
> + index = 2;
> + else if (rc >= RTE_GRAPH_PROFILE_BURST_SMALL)
> + index = 1;
> + else
> + index = 0;
> +
> + node->usage_stats[index].calls++;
> + node->usage_stats[index].cycles += cycles;
> + node->usage_stats[index].objs += rc;
> + }
> +#endif
> } else {
> node->process(graph, node, objs, node->idx);
> }
> --
> 2.43.0
More information about the dev
mailing list