| Message ID | 1701970793-6865-1-git-send-email-si-wei.liu@oracle.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | vdpa-net: improve migration downtime through descriptor ASID and persistent IOTLB | expand |
On Thu, Dec 7, 2023 at 7:50 PM Si-Wei Liu <si-wei.liu@oracle.com> wrote: > > This patch series contain several enhancements to SVQ live migration downtime > for vDPA-net hardware device, specifically on mlx5_vdpa. Currently it is based > off of Eugenio's RFC v2 .load_setup series [1] to utilize the shared facility > and reduce frictions in merging or duplicating code if at all possible. > > It's stacked up in particular order as below, as the optimization for one on > the top has to depend on others on the bottom. Here's a breakdown for what > each part does respectively: > > Patch # | Feature / optimization > ---------V------------------------------------------------------------------- > 35 - 40 | trace events > 34 | migrate_cancel bug fix > 21 - 33 | (Un)map batching at stop-n-copy to further optimize LM down time > 11 - 20 | persistent IOTLB [3] to improve LM down time > 02 - 10 | SVQ descriptor ASID [2] to optimize SVQ switching > 01 | dependent linux headers > V > Hi Si-Wei, Thanks for the series, I think it contains great additions for the live migration solution! It is pretty large though. Do you think it would be feasible to split out the fixes and the tracing patches in a separated series? That would allow the reviews to focus on the downtime reduction. I think I acked all of them. Maybe we can even create a third series for the vring asid? I think we should note an increase on performance using svq so it justifies by itself too. > Let's first define 2 sources of downtime that this work is concerned with: > > * SVQ switching downtime (Downtime #1): downtime at the start of migration. > Time spent on teardown and setup for SVQ mode switching, and this downtime > is regarded as the maxium time for an individual vdpa-net device. > No memory transfer is involved during SVQ switching, hence no . > > * LM downtime (Downtime #2): aggregated downtime for all vdpa-net devices on > resource teardown and setup in the last stop-n-copy phase on source host. > > With each part of the optimizations applied bottom up, the effective outcome > in terms of down time (in seconds) performance can be observed in this table: > > > | Downtime #1 | Downtime #2 > --------------------+-------------------+------------------- > Baseline QEMU | 20s ~ 30s | 20s > | | > Iterative map | | > at destination[1] | 5s | 20s > | | > SVQ descriptor | | > ASID [2] | 2s | 5s > | | > | | > persistent IOTLB | 2s | 2s > [3] | | > | | > (Un)map batching | | > at stop-n-copy | 1.7s | 1.5s > before switchover | | > > (VM config: 128GB mem, 2 mlx5_vdpa devices, each w/ 4 data vqs) > Thanks for all the profiling, it looks promising! > Please find the details regarding each enhancement on the commit log. > > Thanks, > -Siwei > > > [1] [RFC PATCH v2 00/10] Map memory at destination .load_setup in vDPA-net migration > https://lists.nongnu.org/archive/html/qemu-devel/2023-11/msg05711.html > [2] VHOST_BACKEND_F_DESC_ASID > https://lore.kernel.org/virtualization/20231018171456.1624030-2-dtatulea@nvidia.com/ > [3] VHOST_BACKEND_F_IOTLB_PERSIST > https://lore.kernel.org/virtualization/1698304480-18463-1-git-send-email-si-wei.liu@oracle.com/ > > --- > > Si-Wei Liu (40): > linux-headers: add vhost_types.h and vhost.h > vdpa: add vhost_vdpa_get_vring_desc_group > vdpa: probe descriptor group index for data vqs > vdpa: piggyback desc_group index when probing isolated cvq > vdpa: populate desc_group from net_vhost_vdpa_init > vhost: make svq work with gpa without iova translation > vdpa: move around vhost_vdpa_set_address_space_id > vdpa: add back vhost_vdpa_net_first_nc_vdpa > vdpa: no repeat setting shadow_data > vdpa: assign svq descriptors a separate ASID when possible > vdpa: factor out vhost_vdpa_last_dev > vdpa: check map_thread_enabled before join maps thread > vdpa: ref counting VhostVDPAShared > vdpa: convert iova_tree to ref count based > vdpa: add svq_switching and flush_map to header > vdpa: indicate SVQ switching via flag > vdpa: judge if map can be kept across reset > vdpa: unregister listener on last dev cleanup > vdpa: should avoid map flushing with persistent iotlb > vdpa: avoid mapping flush across reset > vdpa: vhost_vdpa_dma_batch_end_once rename > vdpa: factor out vhost_vdpa_map_batch_begin > vdpa: vhost_vdpa_dma_batch_begin_once rename > vdpa: factor out vhost_vdpa_dma_batch_end > vdpa: add asid to dma_batch_once API > vdpa: return int for dma_batch_once API > vdpa: add asid to all dma_batch call sites > vdpa: support iotlb_batch_asid > vdpa: expose API vhost_vdpa_dma_batch_once > vdpa: batch map/unmap op per svq pair basis > vdpa: batch map and unmap around cvq svq start/stop > vdpa: factor out vhost_vdpa_net_get_nc_vdpa > vdpa: batch multiple dma_unmap to a single call for vm stop > vdpa: fix network breakage after cancelling migration > vdpa: add vhost_vdpa_set_address_space_id trace > vdpa: add vhost_vdpa_get_vring_base trace for svq mode > vdpa: add vhost_vdpa_set_dev_vring_base trace for svq mode > vdpa: add trace events for eval_flush > vdpa: add trace events for vhost_vdpa_net_load_cmd > vdpa: add trace event for vhost_vdpa_net_load_mq > > hw/virtio/trace-events | 9 +- > hw/virtio/vhost-shadow-virtqueue.c | 35 ++- > hw/virtio/vhost-vdpa.c | 156 +++++++--- > include/hw/virtio/vhost-vdpa.h | 16 + > include/standard-headers/linux/vhost_types.h | 13 + > linux-headers/linux/vhost.h | 9 + > net/trace-events | 8 + > net/vhost-vdpa.c | 434 ++++++++++++++++++++++----- > 8 files changed, 558 insertions(+), 122 deletions(-) > > -- > 1.8.3.1 >
On Fri, Dec 8, 2023 at 2:50 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote: > > This patch series contain several enhancements to SVQ live migration downtime > for vDPA-net hardware device, specifically on mlx5_vdpa. Currently it is based > off of Eugenio's RFC v2 .load_setup series [1] to utilize the shared facility > and reduce frictions in merging or duplicating code if at all possible. > > It's stacked up in particular order as below, as the optimization for one on > the top has to depend on others on the bottom. Here's a breakdown for what > each part does respectively: > > Patch # | Feature / optimization > ---------V------------------------------------------------------------------- > 35 - 40 | trace events > 34 | migrate_cancel bug fix > 21 - 33 | (Un)map batching at stop-n-copy to further optimize LM down time > 11 - 20 | persistent IOTLB [3] to improve LM down time > 02 - 10 | SVQ descriptor ASID [2] to optimize SVQ switching > 01 | dependent linux headers > V > > Let's first define 2 sources of downtime that this work is concerned with: > > * SVQ switching downtime (Downtime #1): downtime at the start of migration. > Time spent on teardown and setup for SVQ mode switching, and this downtime > is regarded as the maxium time for an individual vdpa-net device. > No memory transfer is involved during SVQ switching, hence no . > > * LM downtime (Downtime #2): aggregated downtime for all vdpa-net devices on > resource teardown and setup in the last stop-n-copy phase on source host. > > With each part of the optimizations applied bottom up, the effective outcome > in terms of down time (in seconds) performance can be observed in this table: > > > | Downtime #1 | Downtime #2 > --------------------+-------------------+------------------- > Baseline QEMU | 20s ~ 30s | 20s > | | > Iterative map | | > at destination[1] | 5s | 20s > | | > SVQ descriptor | | > ASID [2] | 2s | 5s > | | > | | > persistent IOTLB | 2s | 2s > [3] | | > | | > (Un)map batching | | > at stop-n-copy | 1.7s | 1.5s > before switchover | | > > (VM config: 128GB mem, 2 mlx5_vdpa devices, each w/ 4 data vqs) This looks promising! But the series looks a little bit huge, can we split them into 2 or 3 series? It helps to speed up the reviewing and merging. Thanks > > Please find the details regarding each enhancement on the commit log. > > Thanks, > -Siwei > > > [1] [RFC PATCH v2 00/10] Map memory at destination .load_setup in vDPA-net migration > https://lists.nongnu.org/archive/html/qemu-devel/2023-11/msg05711.html > [2] VHOST_BACKEND_F_DESC_ASID > https://lore.kernel.org/virtualization/20231018171456.1624030-2-dtatulea@nvidia.com/ > [3] VHOST_BACKEND_F_IOTLB_PERSIST > https://lore.kernel.org/virtualization/1698304480-18463-1-git-send-email-si-wei.liu@oracle.com/ > > --- > > Si-Wei Liu (40): > linux-headers: add vhost_types.h and vhost.h > vdpa: add vhost_vdpa_get_vring_desc_group > vdpa: probe descriptor group index for data vqs > vdpa: piggyback desc_group index when probing isolated cvq > vdpa: populate desc_group from net_vhost_vdpa_init > vhost: make svq work with gpa without iova translation > vdpa: move around vhost_vdpa_set_address_space_id > vdpa: add back vhost_vdpa_net_first_nc_vdpa > vdpa: no repeat setting shadow_data > vdpa: assign svq descriptors a separate ASID when possible > vdpa: factor out vhost_vdpa_last_dev > vdpa: check map_thread_enabled before join maps thread > vdpa: ref counting VhostVDPAShared > vdpa: convert iova_tree to ref count based > vdpa: add svq_switching and flush_map to header > vdpa: indicate SVQ switching via flag > vdpa: judge if map can be kept across reset > vdpa: unregister listener on last dev cleanup > vdpa: should avoid map flushing with persistent iotlb > vdpa: avoid mapping flush across reset > vdpa: vhost_vdpa_dma_batch_end_once rename > vdpa: factor out vhost_vdpa_map_batch_begin > vdpa: vhost_vdpa_dma_batch_begin_once rename > vdpa: factor out vhost_vdpa_dma_batch_end > vdpa: add asid to dma_batch_once API > vdpa: return int for dma_batch_once API > vdpa: add asid to all dma_batch call sites > vdpa: support iotlb_batch_asid > vdpa: expose API vhost_vdpa_dma_batch_once > vdpa: batch map/unmap op per svq pair basis > vdpa: batch map and unmap around cvq svq start/stop > vdpa: factor out vhost_vdpa_net_get_nc_vdpa > vdpa: batch multiple dma_unmap to a single call for vm stop > vdpa: fix network breakage after cancelling migration > vdpa: add vhost_vdpa_set_address_space_id trace > vdpa: add vhost_vdpa_get_vring_base trace for svq mode > vdpa: add vhost_vdpa_set_dev_vring_base trace for svq mode > vdpa: add trace events for eval_flush > vdpa: add trace events for vhost_vdpa_net_load_cmd > vdpa: add trace event for vhost_vdpa_net_load_mq > > hw/virtio/trace-events | 9 +- > hw/virtio/vhost-shadow-virtqueue.c | 35 ++- > hw/virtio/vhost-vdpa.c | 156 +++++++--- > include/hw/virtio/vhost-vdpa.h | 16 + > include/standard-headers/linux/vhost_types.h | 13 + > linux-headers/linux/vhost.h | 9 + > net/trace-events | 8 + > net/vhost-vdpa.c | 434 ++++++++++++++++++++++----- > 8 files changed, 558 insertions(+), 122 deletions(-) > > -- > 1.8.3.1 >
This patch series contain several enhancements to SVQ live migration downtime for vDPA-net hardware device, specifically on mlx5_vdpa. Currently it is based off of Eugenio's RFC v2 .load_setup series [1] to utilize the shared facility and reduce frictions in merging or duplicating code if at all possible. It's stacked up in particular order as below, as the optimization for one on the top has to depend on others on the bottom. Here's a breakdown for what each part does respectively: Patch # | Feature / optimization ---------V------------------------------------------------------------------- 35 - 40 | trace events 34 | migrate_cancel bug fix 21 - 33 | (Un)map batching at stop-n-copy to further optimize LM down time 11 - 20 | persistent IOTLB [3] to improve LM down time 02 - 10 | SVQ descriptor ASID [2] to optimize SVQ switching 01 | dependent linux headers V Let's first define 2 sources of downtime that this work is concerned with: * SVQ switching downtime (Downtime #1): downtime at the start of migration. Time spent on teardown and setup for SVQ mode switching, and this downtime is regarded as the maxium time for an individual vdpa-net device. No memory transfer is involved during SVQ switching, hence no . * LM downtime (Downtime #2): aggregated downtime for all vdpa-net devices on resource teardown and setup in the last stop-n-copy phase on source host. With each part of the optimizations applied bottom up, the effective outcome in terms of down time (in seconds) performance can be observed in this table: | Downtime #1 | Downtime #2 --------------------+-------------------+------------------- Baseline QEMU | 20s ~ 30s | 20s | | Iterative map | | at destination[1] | 5s | 20s | | SVQ descriptor | | ASID [2] | 2s | 5s | | | | persistent IOTLB | 2s | 2s [3] | | | | (Un)map batching | | at stop-n-copy | 1.7s | 1.5s before switchover | | (VM config: 128GB mem, 2 mlx5_vdpa devices, each w/ 4 data vqs) Please find the details regarding each enhancement on the commit log. Thanks, -Siwei [1] [RFC PATCH v2 00/10] Map memory at destination .load_setup in vDPA-net migration https://lists.nongnu.org/archive/html/qemu-devel/2023-11/msg05711.html [2] VHOST_BACKEND_F_DESC_ASID https://lore.kernel.org/virtualization/20231018171456.1624030-2-dtatulea@nvidia.com/ [3] VHOST_BACKEND_F_IOTLB_PERSIST https://lore.kernel.org/virtualization/1698304480-18463-1-git-send-email-si-wei.liu@oracle.com/ --- Si-Wei Liu (40): linux-headers: add vhost_types.h and vhost.h vdpa: add vhost_vdpa_get_vring_desc_group vdpa: probe descriptor group index for data vqs vdpa: piggyback desc_group index when probing isolated cvq vdpa: populate desc_group from net_vhost_vdpa_init vhost: make svq work with gpa without iova translation vdpa: move around vhost_vdpa_set_address_space_id vdpa: add back vhost_vdpa_net_first_nc_vdpa vdpa: no repeat setting shadow_data vdpa: assign svq descriptors a separate ASID when possible vdpa: factor out vhost_vdpa_last_dev vdpa: check map_thread_enabled before join maps thread vdpa: ref counting VhostVDPAShared vdpa: convert iova_tree to ref count based vdpa: add svq_switching and flush_map to header vdpa: indicate SVQ switching via flag vdpa: judge if map can be kept across reset vdpa: unregister listener on last dev cleanup vdpa: should avoid map flushing with persistent iotlb vdpa: avoid mapping flush across reset vdpa: vhost_vdpa_dma_batch_end_once rename vdpa: factor out vhost_vdpa_map_batch_begin vdpa: vhost_vdpa_dma_batch_begin_once rename vdpa: factor out vhost_vdpa_dma_batch_end vdpa: add asid to dma_batch_once API vdpa: return int for dma_batch_once API vdpa: add asid to all dma_batch call sites vdpa: support iotlb_batch_asid vdpa: expose API vhost_vdpa_dma_batch_once vdpa: batch map/unmap op per svq pair basis vdpa: batch map and unmap around cvq svq start/stop vdpa: factor out vhost_vdpa_net_get_nc_vdpa vdpa: batch multiple dma_unmap to a single call for vm stop vdpa: fix network breakage after cancelling migration vdpa: add vhost_vdpa_set_address_space_id trace vdpa: add vhost_vdpa_get_vring_base trace for svq mode vdpa: add vhost_vdpa_set_dev_vring_base trace for svq mode vdpa: add trace events for eval_flush vdpa: add trace events for vhost_vdpa_net_load_cmd vdpa: add trace event for vhost_vdpa_net_load_mq hw/virtio/trace-events | 9 +- hw/virtio/vhost-shadow-virtqueue.c | 35 ++- hw/virtio/vhost-vdpa.c | 156 +++++++--- include/hw/virtio/vhost-vdpa.h | 16 + include/standard-headers/linux/vhost_types.h | 13 + linux-headers/linux/vhost.h | 9 + net/trace-events | 8 + net/vhost-vdpa.c | 434 ++++++++++++++++++++++----- 8 files changed, 558 insertions(+), 122 deletions(-)