hi team,
recently I am doing plugin dev with VPP, I found all threads are spending too
many CPU cycles with locking/unlocking (with one numa node)buffer_main's
buffer_known_hash_lockp, it hinders burst performance.
```
+ 6.00% vpp_wk_4 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 5.33% vpp_wk_7 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 5.29% vpp_wk_9 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 5.18% vpp_wk_2 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 5.05% vpp_wk_0 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.95% vpp_wk_10 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.92% vpp_wk_6 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.90% vpp_wk_3 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.63% vpp_wk_8 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.41% vpp_wk_5 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.33% vpp_wk_1 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 4.24% vpp_wk_11 libvlib.so.22.02.0 [.] clib_spinlock_lock
+ 2.27% vpp_main libvlib.so.22.02.0 [.] dispatch_node
+ 0.89% vpp_main libvlib.so.22.02.0 [.]
vlib_main_or_worker_loop
+ 0.85% vpp_main libvlib.so.22.02.0 [.] vlib_get_node
+ 0.83% vpp_wk_11 libvlib.so.22.02.0 [.] vlib_frame_free
+ 0.78% vpp_wk_1 libvlib.so.22.02.0 [.] vlib_frame_free
+ 0.65% vpp_wk_5 libvlib.so.22.02.0 [.] vlib_frame_free
+ 0.51% vpp_main vcgnat_dp_plugin.so [.]
node_channel_tx_process
```
the calling graph:
- 5.33% vpp_wk_7 libvlib.so.22.02.0 [.] clib_spinlock_lock
- clib_spinlock_lock
- 4.89% vlib_buffer_is_known
- vlib_buffer_validate_alloc_free
+ 1.94% vlib_buffer_alloc_from_pool
+ 1.45% vlib_buffer_pool_put
+ 1.42% vlib_buffer_alloc_from_pool
- 5.29% vpp_wk_9 libvlib.so.22.02.0 [.] clib_spinlock_lock
- clib_spinlock_lock
- 4.89% vlib_buffer_is_known
- vlib_buffer_validate_alloc_free
+ 2.14% vlib_buffer_alloc_from_pool
+ 1.40% vlib_buffer_alloc_from_pool
+ 1.28% vlib_buffer_pool_put
+ 5.18% vpp_wk_2 libvlib.so.22.02.0 [.] clib_spinlock_lock
- 5.05% vpp_wk_0 libvlib.so.22.02.0 [.] clib_spinlock_lock
- clib_spinlock_lock
- 4.57% vlib_buffer_is_known
- vlib_buffer_validate_alloc_free
+ 2.25% vlib_buffer_alloc_from_pool
+ 1.20% vlib_buffer_pool_put
+ 1.03% vlib_buffer_alloc_from_pool
thanks,
Looking forward to your help.
-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#21810): https://lists.fd.io/g/vpp-dev/message/21810
Mute This Topic: https://lists.fd.io/mt/93199622/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/leave/1480452/21656/631435203/xyzzy
[arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-
