Re: [PATCH v7 0/7] Stage-Ordered API and other extensions for ring library

[Stephen Hemminger <stephen@networkplumber.org>]

On Wed, 30 Oct 2024 17:22:57 -0400
Konstantin Ananyev <konstantin.ananyev@huawei.com> wrote:

> Testing coverage (passed):
> x86_64, i686, PPC, ARM   
> 
> Would like to express my gratitude to all community members who helped me
> with testing it on different platforms, in particular:
> David Christensen <drc@linux.ibm.com>
> Cody Cheng <ccheng@iol.unh.edu>
> Patrick Robb <probb@iol.unh.edu>
> Phanendra Vukkisala <pvukkisala@marvell.com>
> Chengwen Feng <fengchengwen@huawei.com>
> 
> v6 -> v7
> - updated Programmer Guide (Jerin, Morten, Stephen)
> - fix some functions in public headers without comments (Morten)
> - update debug checks, added new macro for that: RTE_SORING_DEBUG
>   (disabled by default).
> 
> v5 -> v6
> - fix problem with ring_stress_autotest (Phanendra)
> - added more checks and debug output
> 
> v4 -> v5
> - fix public API/doc comments from Jerin
> - update devtools/build-dict.sh (Stephen)
> - fix MSVC warnings
> - introduce new test-suite for meson (stress) with
>   ring_stress_autotest and soring_stress_autotest in it
> - enhance error report in tests
> - reorder some sync code in soring and add extra checks
>   (for better debuggability)
> 
> v3 -> v4:
> - fix compilation/doxygen complains (attempt #2)
> - updated release notes
> 
> v2 -> v3:
> - fix compilation/doxygen complains
> - dropped patch:
>   "examples/l3fwd: make ACL work in pipeline and eventdev modes": [2]
>   As was mentioned in the patch desctiption it was way too big,
>   controversial and incomplete. If the community is ok to introduce
>   pipeline model into the l3fwd, then it is propbably worth to be
>   a separate patch series.
> 
> v1 -> v2:
> - rename 'elmst/objst' to 'meta' (Morten)
> - introduce new data-path APIs set: one with both meta{} and objs[],
>   second with just objs[] (Morten)
> - split data-path APIs into burst/bulk flavours (same as rte_ring)
> - added dump function for te_soring and improved dump() for rte_ring.
> - dropped patch:
>   " ring: minimize reads of the counterpart cache-line"
>   - no performance gain observed
>   - actually it does change behavior of conventional rte_ring
>     enqueue/dequeue APIs -
>     it could return available/free less then actually exist in the ring.
>     As in some other libs we reliy on that information - it will
>     introduce problems.
> 
> The main aim of these series is to extend ring library with
> new API that allows user to create/use Staged-Ordered-Ring (SORING)
> abstraction. In addition to that there are few other patches that serve
> different purposes:
> - first two patches are just code reordering to de-duplicate
>   and generalize existing rte_ring code.
> - patch #3 extends rte_ring_dump() to correctly print head/tail metadata
>   for different sync modes.
> - next two patches introduce SORING API into the ring library and
>   provide UT for it.
> 
> SORING overview
> ===============
> Staged-Ordered-Ring (SORING) provides a SW abstraction for 'ordered' queues
> with multiple processing 'stages'. It is based on conventional DPDK
> rte_ring, re-uses many of its concepts, and even substantial part of
> its code.
> It can be viewed as an 'extension' of rte_ring functionality.
> In particular, main SORING properties:
> - circular ring buffer with fixed size objects
> - producer, consumer plus multiple processing stages in between.
> - allows to split objects processing into multiple stages.
> - objects remain in the same ring while moving from one stage to the other,
>   initial order is preserved, no extra copying needed.
> - preserves the ingress order of objects within the queue across multiple
>   stages
> - each stage (and producer/consumer) can be served by single and/or
>   multiple threads.
> 
> - number of stages, size and number of objects in the ring are
>  configurable at ring initialization time.
> 
> Data-path API provides four main operations:
> - enqueue/dequeue works in the same manner as for conventional rte_ring,
>   all rte_ring synchronization types are supported.
> - acquire/release - for each stage there is an acquire (start) and
>   release (finish) operation. After some objects are 'acquired' -
>   given thread can safely assume that it has exclusive ownership of
>   these objects till it will invoke 'release' for them.
>   After 'release', objects can be 'acquired' by next stage and/or dequeued
>   by the consumer (in case of last stage).
> 
> Expected use-case: applications that uses pipeline model
> (probably with multiple stages) for packet processing, when preserving
> incoming packet order is important.
> 
> The concept of ‘ring with stages’ is similar to DPDK OPDL eventdev PMD [1],
> but the internals are different.
> In particular, SORING maintains internal array of 'states' for each element
> in the ring that is  shared by all threads/processes that access the ring.
> That allows 'release' to avoid excessive waits on the tail value and helps
> to improve performancei and scalability.
> In terms of performance, with our measurements rte_soring and
> conventional rte_ring provide nearly identical numbers.
> As an example, on our SUT: Intel ICX CPU @ 2.00GHz,
> l3fwd (--lookup=acl) in pipeline mode [2] both
> rte_ring and rte_soring reach ~20Mpps for single I/O lcore and same
> number of worker lcores.
> 
> [1] https://www.dpdk.org/wp-content/uploads/sites/35/2018/06/DPDK-China2017-Ma-OPDL.pdf
> [2] https://patchwork.dpdk.org/project/dpdk/patch/20240906131348.804-7-konstantin.v.ananyev@yandex.ru/

One future suggestion. What about having an example (l3fwd-soring?) so
that performance can be compared.

Assuming you get the other minor comments from Morten fixed.

Series-Acked-by: Stephen Hemminger <stephen@networkplumber.org>