[PATCH v15 00/16] graph enhancement for multi-core dispatch

[David Marchand]

On Wed, Jun 14, 2023 at 6:13 PM Zhirun Yan <zhirun.yan at intel.com> wrote:
>
> V15:
> Fix build issues. (add unused macro in patch 10 then remove in patch 11,
> fix log message format in patch 01, rm wrong assert in dispatch_walk in patch 12)
>
> V14:
> Rebase to main and fix build issues.(align header name in patch 01,02)
>
> V13:
> Add sub header into meson indirect_headers list to export.(change meson.build in patch02, 05)
>
> V12:
> Fix compilation broken at patch 1.(keep renamed header align with patch 1,2)
>
> V11:
> Update comments and fix to add experimental flags for rte_graph_model_is_valid() in patch 04.
> Update added symbols in alphabetical order in version.map with patch 04,05,06,08,10.
> Update commit message in patch 16.
>
> V10:
> Add rte_graph_worker_model_no_check_get() for fast path, extract rte_graph_model_is_valid()
> in patch 04.
> Change RTE_ASSERT to return in patch 06.
> Change to treat not defined RTE_GRAPH_MODEL_SELECT as runtime pick in patch 13.
> Move stats into dispatch union in patch 14.
> Change example to align with RTE_GRAPH_MODEL_SELECT scheme in patch 16.
> Squash patch 17(doc) into patch 13(prog_guide), 16(example guide).
>
>
> V9:
> Fix CI build issues for doc building(move TAILQ next pointer out of union) in patch 09,10.
> Fix graph model check in rte_graph_worker_model_set() in patch 04.
> Fix typo in doc.
>
> V8:
> No performance dorp for original l3fwd-graph and graph_perf_autotest.
>
> Update graph model set/get functions and add graph_model_is_valid() in patch 04.
> Update doc for new scheme usage(choose model in runtime or compile time).
> Update dispatch schedule struct into union.
> Change enum rte_graph_worker_model to macro define in rte_graph_worker_common.h.
> Add model clone in graph_clone() in patch 08.
> Remove unnecessary inline for slow path func graph_src_node_avail() in patch 06.
>
>
> V7:
> Revert rte_rdtsc_precise() in fastpath to fix performance issues in patch 03.
> Introduce new scheme for model choosing. Use RTE_GRAPH_MODEL_SELECT to choose in
>   compile-time in patch 13, 15.(must have rte_graph_worker_model_set() to help
>   other config func to do model specific things like alloc wq, collect stats)
> Extract the common func clone_name() into graph_private.h for graph/node clone in
>   patch 07.(new patch)
> Use rte_graph->model in rte_graph_worker_model_set() instead of RTE_PER_LCORE_*.
> Add test case for all new APIs in patch 16(new patch).
> Remove *_END line in enum rte_graph_worker_model in patch 04.
> Add model check for graph lcore binding.
> Rename workqueue as graph_mcore_dispatch_wq_node in patch 09.
> Change all new model files/APIs with prefix _mcore_dispatch_.
> Change description of new API, comments of func/structure to explicitly mention for
>   mcore dispatch model only. Add Doxygen comments.
> Update l3fwd-graph with new scheme, Update doc.
> Update MAINTAINERS.
> Fix typo and format issues.
>
> V6:
> Change rte_rdtsc() to rte_rdtsc_precise().
> Add union in rte_graph_param to configure models.
> Remove memset in fastpath, add RTE_ASSERT for cloned graph.
> Update copyright in patch 02.
> Update l3fwd-graph node affinity, start from rx core successively.
>
> V5:
> Fix CI build issues about dynamically update doc.
>
> V4:
> Fix CI build issues about undefined reference of sched apis.
> Remove inline for model setting.
>
> V3:
> Fix CI build issues about TLS and typo.
>
> V2:
> Use git mv to keep git history.
> Use TLS for per-thread local storage.
> Change model name to mcore dispatch.
> Change API with specific mode name.
> Split big patch.
> Fix CI issues.
> Rebase l3fwd-graph example.
> Update doc and maintainers files.
>
>
> Currently, rte_graph supports RTC (Run-To-Completion) model within each
> of a single core.
> RTC is one of the typical model of packet processing. Others like
> Pipeline or Hybrid are lack of support.
>
> The patch set introduces a 'multicore dispatch' model selection which
> is a self-reacting scheme according to the core affinity.
> The new model enables a cross-core dispatching mechanism which employs a
> scheduling work-queue to dispatch streams to other worker cores which
> being associated with the destination node. When core flavor of the
> destination node is a default 'current', the stream can be continue
> executed as normal.
>
> Example:
> 3-node graph targets 3-core budget
>
> RTC:
> Graph: node-0 -> node-1 -> node-2 @Core0.
>
> + - - - - - - - - - - - - - - - - - - - - - +
> '                Core #0/1/2                '
> '                                           '
> ' +--------+     +---------+     +--------+ '
> ' | Node-0 | --> | Node-1  | --> | Node-2 | '
> ' +--------+     +---------+     +--------+ '
> '                                           '
> + - - - - - - - - - - - - - - - - - - - - - +
>
> Dispatch:
>
> Graph topo: node-0 -> Core1; node-1 -> node-2; node-2 -> node-3.
> Config graph: node-0 @Core0; node-1/3 @Core1; node-2 @Core2.
>
> .. code-block:: diff
>
>     + - - - - - -+     +- - - - - - - - - - - - - +     + - - - - - -+
>     '  Core #0   '     '          Core #1         '     '  Core #2   '
>     '            '     '                          '     '            '
>     ' +--------+ '     ' +--------+    +--------+ '     ' +--------+ '
>     ' | Node-0 | - - - ->| Node-1 |    | Node-3 |<- - - - | Node-2 | '
>     ' +--------+ '     ' +--------+    +--------+ '     ' +--------+ '
>     '            '     '     |                    '     '      ^     '
>     + - - - - - -+     +- - -|- - - - - - - - - - +     + - - -|- - -+
>                              |                                 |
>                              + - - - - - - - - - - - - - - - - +
>
>
> The patch set has been break down as below:
>
> 1. Split graph worker into common and default model part.
> 2. Inline graph node processing to make it reusable.
> 3. Add set/get APIs to choose worker model.
> 4. Introduce core affinity API to set the node run on specific worker core.
>   (only use in new model)
> 5. Introduce graph affinity API to bind one graph with specific worker
>   core.
> 6. Introduce graph clone API.
> 7. Introduce stream moving with scheduler work-queue in patch 8~12.
> 8. Add stats for new models.
> 9. Abstract default graph config process and integrate new model into
>   example/l3fwd-graph. Add new parameters for model choosing.
>
> We could run with new worker model by this:
> ./dpdk-l3fwd-graph -l 8,9,10,11 -n 4 -- -p 0x1 --config="(0,0,9)" -P
> --model="dispatch"
>
> References:
> https://static.sched.com/hosted_files/dpdkuserspace22/a6/graph%20introduce%20remote%20dispatch%20for%20mult-core%20scaling.pdf
>
>
> Zhirun Yan (16):
>   graph: rename rte graph worker header as common
>   graph: split graph worker into common and default model
>   graph: move node process into inline function
>   graph: add get/set graph worker model APIs
>   graph: introduce graph node core affinity API
>   graph: introduce graph bind unbind API
>   graph: move node clone name func into private as common
>   graph: introduce graph clone API for other worker core
>   graph: add structure for stream moving between cores
>   graph: introduce stream moving cross cores
>   graph: enable create and destroy graph scheduling workqueue
>   graph: introduce graph walk by cross-core dispatch
>   graph: enable graph multicore dispatch scheduler model
>   graph: add stats for mcore dispatch model
>   test/graph: add functional tests for mcore dispatch model
>   examples/l3fwd-graph: introduce mcore dispatch worker model

Series applied, thanks.


-- 
David Marchand