Re: [dpdk-dev] [PATCH 0/5] vhost: I-cache pressure optimizations

[David Marchand <david.marchand@redhat.com>]

On Fri, May 17, 2019 at 2:23 PM Maxime Coquelin <maxime.coquelin@redhat.com>
wrote:

> Some OVS-DPDK PVP benchmarks show a performance drop
> when switching from DPDK v17.11 to v18.11.
>
> With the addition of packed ring layout support,
> rte_vhost_enqueue_burst and rte_vhost_dequeue_burst
> became very large, and only a part of the instructions
> are executed (either packed or split ring used).
>
> This series aims at improving the I-cache pressure,
> first by un-inlining split and packed rings, but
> also by moving parts considered as cold in dedicated
> functions (dirty page logging, fragmented descriptors
> buffer management added for CVE-2018-1059).
>
> With the series applied, size of the enqueue and
> dequeue split paths is reduced significantly:
>
> +---------+--------------------+---------------------+
> | Version | Enqueue split path |  Dequeue split path |
> +---------+--------------------+---------------------+
> | v19.05  | 16461B             | 25521B              |
> | +series | 7286B              | 11285B              |
> +---------+--------------------+---------------------+
>
> Using perf tool to monitor iTLB-load-misses event
> while doing PVP benchmark with testpmd as vswitch,
> we can see the number of iTLB misses being reduced:
>
> - v19.05:
> # perf stat --repeat 10  -C 2,3  -e iTLB-load-miss -- sleep 10
>
>  Performance counter stats for 'CPU(s) 2,3' (10 runs):
>
>              2,438      iTLB-load-miss
>             ( +- 13.43% )
>
>        10.00058928 +- 0.00000336 seconds time elapsed  ( +-  0.00% )
>
> - +series:
> # perf stat --repeat 10  -C 2,3  -e iTLB-load-miss -- sleep 10
>
>  Performance counter stats for 'CPU(s) 2,3' (10 runs):
>
>                 55      iTLB-load-miss
>             ( +- 10.08% )
>
>        10.00059466 +- 0.00000283 seconds time elapsed  ( +-  0.00% )
>
> The series also force the inlining of some rte_memcpy
> helpers, as by adding packed ring support, some of them
> were not more inlined but embedded as functions in
> the virtio_net object file, which was not expected.
>
> Finally, the series simplifies the descriptors buffers
> prefetching, by doing it in the recently introduced
> descriptor buffer mapping function.
>
> Maxime Coquelin (4):
>   vhost: un-inline dirty pages logging functions
>   vhost: do not inline packed and split functions
>   vhost: do not inline unlikely fragmented buffers code
>   vhost: simplify descriptor's buffer prefetching
>
> root (1):
>   eal/x86: force inlining of all memcpy and mov helpers
>

root ? "oops" :-)


-- 
David Marchand