Optimized scheduler priority queue and packet IO poll queue implementation
with a ring based data structure. Scheduler performance is improved tens of
percents in a typical test case. Also there's no more dependency to pools,
buffer and queue APIs.
l2fwd packet rate, odp-linux with dpdk, 2 x 10GE
------------------------------------------------
cores
mode 1 2 4 8
parallel +24,1% +47,2% +48,1% +53,2%
atomic +20,7% +35,9% +37,5% +44,1%
ordered +9,2% +18,2% +38,2% +39,5%
execution time of odp_scheduling:
---------------------------------
cores
4 8 10 20
master 15,3s 62,5s 109s 884,9s
patch 13s 47s 74s 310s
speed up 1,18 1,33 1,47 2,85
Petri Savolainen (6):
linux-gen: sched: ring based priority queues
linux-gen: sched: ring based pktio command queues
linux-gen: sched: increase number of pktin
linux-gen: sched: pktio poll loop optimization
linux-gen: sched: increase max dequeue batch size
linux-gen: sched: call schedule_release_atomic directly
.../linux-generic/include/odp_schedule_internal.h | 5 +-
platform/linux-generic/odp_schedule.c | 577 +++++++++++----------
2 files changed, 314 insertions(+), 268 deletions(-)
--
2.8.1
