On 12/20/18 6:27 PM, Maxime Coquelin wrote:
First version of this series did merge out-of-order mergeable and non-mergeable receive paths, but Intel STV team highlighted some performance regression when using multiqueue with two cores enqueueing descs on host, while a single core dequeues the two queues. I didn't manage to close the performance gap, so I decided to give-up this refactoring. But while trying to optimize, I reworked the meargeable function so that it looks like the in-order one. I.e. descriptors are now dequeued in batches, so are descriptors refilled. Doing that, I measure a perfromance gain of 6% when doing rxonly microbenchmark with two cores on host, one in guest. Changes since v2: - Rebase after Jens' packed ring series - Break refill loop if VQ_RING_DESC_CHAIN_END (Tiwei) - Fixup commit message Maxime Coquelin (3): net/virtio: inline refill and offload helpers net/virtio: add non-mergeable support to in-order path net/virtio: improve batching in mergeable path drivers/net/virtio/virtio_ethdev.c | 14 +- drivers/net/virtio/virtio_ethdev.h | 2 +- drivers/net/virtio/virtio_rxtx.c | 257 ++++++++++++++++------------- 3 files changed, 145 insertions(+), 128 deletions(-)
Applied to dpdk-next-virtio Maxime
