[dpdk-dev] [PATCH v6 0/6] distributor library performance enhancements

[Bruce Richardson]

On Mon, Jan 23, 2017 at 09:24:34AM +0000, David Hunt wrote:
> This patch aims to improve the throughput of the distributor library.
> 
> It uses a similar handshake mechanism to the previous version of
> the library, in that bits are used to indicate when packets are ready
> to be sent to a worker and ready to be returned from a worker. One main
> difference is that instead of sending one packet in a cache line, it makes
> use of the 7 free spaces in the same cache line in order to send up to
> 8 packets at a time to/from a worker.
> 
> The flow matching algorithm has had significant re-work, and now keeps an
> array of inflight flows and an array of backlog flows, and matches incoming
> flows to the inflight/backlog flows of all workers so that flow pinning to
> workers can be maintained.
> 
> The Flow Match algorithm has both scalar and a vector versions, and a
> function pointer is used to select the post appropriate function at run time,
> depending on the presence of the SSE2 cpu flag. On non-x86 platforms, the
> the scalar match function is selected, which should still gives a good boost
> in performance over the non-burst API.
> 
> v2 changes:
>   * Created a common distributor_priv.h header file with common
>     definitions and structures.
>   * Added a scalar version so it can be built and used on machines without
>     sse2 instruction set
>   * Added unit autotests
>   * Added perf autotest
> 
> v3 changes:
>   * Addressed mailing list review comments
>   * Test code removal
>   * Split out SSE match into separate file to facilitate NEON addition
>   * Cleaned up conditional compilation flags for SSE2
>   * Addressed c99 style compilation errors
>   * rebased on latest head (Jan 2 2017, Happy New Year to all)
> 
> v4 changes:
>    * fixed issue building shared libraries
> 
> v5 changes:
>    * Removed some un-needed code around retries in worker API calls
>    * Cleanup due to review comments on mailing list
>    * Cleanup of non-x86 platform compilation, fallback to scalar match
> 
> v6 changes:
>    * Fixed intermittent segfault where num pkts not divisible
>      by BURST_SIZE
>    * Cleanup due to review comments on mailing list
>    * Renamed _priv.h to _private.h.
> 
> Notes:
>    Apps must now work in bursts, as up to 8 are given to a worker at a time
>    For performance in matching, Flow ID's are 15-bits
>    Original API (and code) is kept for backward compatibility
> 
> Performance Gains
>    2.2GHz Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz
>    2 x XL710 40GbE NICS to 2 x 40Gbps traffic generator channels 64b packets
>    separate cores for rx, tx, distributor
>     1 worker  - 4.8x
>     4 workers - 2.9x
>     8 workers - 1.8x
>    12 workers - 2.1x
>    16 workers - 1.8x
> 

Series Acked-by: Bruce Richardson <bruce.richardson at intel.com>