| Message ID | 20230913200045.1024233-1-stefanha@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | virtio-blk: use blk_io_plug_call() instead of notification BH | expand |
On Wed, Sep 13, 2023 at 04:00:41PM -0400, Stefan Hajnoczi wrote: > v3: > - Add comment pointing to API documentation in .c file [Philippe] > - Add virtio_notify_irqfd_deferred_fn trace event [Ilya] > - Remove outdated #include [Ilya] > v2: > - Rename blk_io_plug() to defer_call() and move it to util/ so the net > subsystem can use it [Ilya] > - Add defer_call_begin()/end() to thread_pool_completion_bh() to match Linux > AIO and io_uring completion batching > > Replace the seldom-used virtio-blk notification BH mechanism with > blk_io_plug(). This is part of an effort to enable the multi-queue block layer > in virtio-blk. The notification BH was not multi-queue friendly. > > The blk_io_plug() mechanism improves fio rw=randread bs=4k iodepth=64 numjobs=8 > IOPS by ~9% with a single IOThread and 8 vCPUs (this is not even a multi-queue > block layer configuration) compared to no completion batching. iodepth=1 > decreases by ~1% but this could be noise. Benchmark details are available here: > https://gitlab.com/stefanha/virt-playbooks/-/tree/blk_io_plug-irqfd virtio things: Reviewed-by: Michael S. Tsirkin <mst@redhat.com> regression is a bit worrysome if real though - just do a bit longer testing to make sure? > Stefan Hajnoczi (4): > block: rename blk_io_plug_call() API to defer_call() > util/defer-call: move defer_call() to util/ > virtio: use defer_call() in virtio_irqfd_notify() > virtio-blk: remove batch notification BH > > MAINTAINERS | 3 +- > include/qemu/defer-call.h | 16 +++ > include/sysemu/block-backend-io.h | 4 - > block/blkio.c | 9 +- > block/io_uring.c | 11 ++- > block/linux-aio.c | 9 +- > block/nvme.c | 5 +- > block/plug.c | 159 ------------------------------ > hw/block/dataplane/virtio-blk.c | 48 +-------- > hw/block/dataplane/xen-block.c | 11 ++- > hw/block/virtio-blk.c | 5 +- > hw/scsi/virtio-scsi.c | 7 +- > hw/virtio/virtio.c | 13 ++- > util/defer-call.c | 156 +++++++++++++++++++++++++++++ > util/thread-pool.c | 5 + > block/meson.build | 1 - > hw/virtio/trace-events | 1 + > util/meson.build | 1 + > 18 files changed, 231 insertions(+), 233 deletions(-) > create mode 100644 include/qemu/defer-call.h > delete mode 100644 block/plug.c > create mode 100644 util/defer-call.c > > -- > 2.41.0
Am 13.09.2023 um 22:00 hat Stefan Hajnoczi geschrieben: > v3: > - Add comment pointing to API documentation in .c file [Philippe] > - Add virtio_notify_irqfd_deferred_fn trace event [Ilya] > - Remove outdated #include [Ilya] > v2: > - Rename blk_io_plug() to defer_call() and move it to util/ so the net > subsystem can use it [Ilya] > - Add defer_call_begin()/end() to thread_pool_completion_bh() to match Linux > AIO and io_uring completion batching > > Replace the seldom-used virtio-blk notification BH mechanism with > blk_io_plug(). This is part of an effort to enable the multi-queue block layer > in virtio-blk. The notification BH was not multi-queue friendly. > > The blk_io_plug() mechanism improves fio rw=randread bs=4k iodepth=64 numjobs=8 > IOPS by ~9% with a single IOThread and 8 vCPUs (this is not even a multi-queue > block layer configuration) compared to no completion batching. iodepth=1 > decreases by ~1% but this could be noise. Benchmark details are available here: > https://gitlab.com/stefanha/virt-playbooks/-/tree/blk_io_plug-irqfd Thanks, applied to the block branch. Kevin