diff options
Diffstat (limited to 'docs/gettingstarted/troubleshooting/cpuusage.rst')
-rw-r--r-- | docs/gettingstarted/troubleshooting/cpuusage.rst | 112 |
1 files changed, 112 insertions, 0 deletions
diff --git a/docs/gettingstarted/troubleshooting/cpuusage.rst b/docs/gettingstarted/troubleshooting/cpuusage.rst new file mode 100644 index 00000000000..9b4514e128e --- /dev/null +++ b/docs/gettingstarted/troubleshooting/cpuusage.rst @@ -0,0 +1,112 @@ +.. _cpuusage: + +************** +CPU Load/Usage +************** + +There are various commands and tools that can help users see FD.io VPP CPU and memory usage at runtime. + +Linux top/htop +============== + +The Linux top and htop are decent tools to look at FD.io VPP cpu and memory usage, but they will only show +preallocated memory and total CPU usage. These commands can be useful to show which cores VPP is running on. + +This is an example of VPP instance that is running on cores 8 and 9. For this output type **top** and then +type **1** when the tool starts. + +.. code-block:: console + + $ top + + top - 11:04:04 up 35 days, 3:16, 5 users, load average: 2.33, 2.23, 2.16 + Tasks: 435 total, 2 running, 432 sleeping, 1 stopped, 0 zombie + %Cpu0 : 1.0 us, 0.7 sy, 0.0 ni, 98.0 id, 0.0 wa, 0.0 hi, 0.3 si, 0.0 st + %Cpu1 : 2.0 us, 0.3 sy, 0.0 ni, 97.7 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu2 : 0.7 us, 1.0 sy, 0.0 ni, 98.3 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu3 : 1.7 us, 0.7 sy, 0.0 ni, 97.7 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu4 : 2.0 us, 0.7 sy, 0.0 ni, 97.4 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu5 : 3.0 us, 0.3 sy, 0.0 ni, 96.7 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu6 : 2.3 us, 0.7 sy, 0.0 ni, 97.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu7 : 2.6 us, 0.3 sy, 0.0 ni, 97.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu8 : 96.0 us, 0.3 sy, 0.0 ni, 3.6 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu9 :100.0 us, 0.0 sy, 0.0 ni, 0.0 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + %Cpu10 : 1.0 us, 0.3 sy, 0.0 ni, 98.7 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st + .... + +VPP Memory Usage +================ + +For details on VPP memory usage you can use the **show memory** command + +This is the example VPP memory usage on 2 cores. + +.. code-block:: console + + # vppctl show memory verbose + Thread 0 vpp_main + 22043 objects, 17878k of 20826k used, 2426k free, 2396k reclaimed, 346k overhead, 1048572k capacity + alloc. from small object cache: 22875 hits 39973 attempts (57.23%) replacements 5143 + alloc. from free-list: 44732 attempts, 26017 hits (58.16%), 528461 considered (per-attempt 11.81) + alloc. from vector-expand: 3430 + allocs: 52324 2027.84 clocks/call + frees: 30280 594.38 clocks/call + Thread 1 vpp_wk_0 + 22043 objects, 17878k of 20826k used, 2427k free, 2396k reclaimed, 346k overhead, 1048572k capacity + alloc. from small object cache: 22881 hits 39984 attempts (57.23%) replacements 5148 + alloc. from free-list: 44736 attempts, 26021 hits (58.17%), 528465 considered (per-attempt 11.81) + alloc. from vector-expand: 3430 + allocs: 52335 2027.54 clocks/call + frees: 30291 594.36 clocks/call + +VPP CPU Load +============ + +To find the VPP CPU load or how busy VPP is use the **show runtime** command. + +With at least one interface in polling mode, the VPP CPU utilization is always 100%. + +A good indicator of CPU load is **"average vectors/node"**. A bigger number means VPP +is more busy but also more efficient. The Maximum value is 255 (unless you change VLIB_FRAME_SIZE in code). +It basically means how many packets are processed in batch. + +If VPP is not loaded it will likely poll so fast that it will just get one or few +packets from the rx queue. This is the case shown below on Thread 1. As load goes up vpp +will have more work to do, so it will poll less frequently, and that will result in more +packets waiting in rx queue. More packets will result in more efficient execution of the +code so number of clock cycles / packet will go down. When "average vectors/node" goes up +close to 255, you will likely start observing rx queue tail drops. + +.. code-block:: console + + # vppctl show run + Thread 0 vpp_main (lcore 8) + Time 6152.9, average vectors/node 0.00, last 128 main loops 0.00 per node 0.00 + vector rates in 0.0000e0, out 0.0000e0, drop 0.0000e0, punt 0.0000e0 + Name State Calls Vectors Suspends Clocks Vectors/Call + acl-plugin-fa-cleaner-process event wait 0 0 1 3.66e4 0.00 + admin-up-down-process event wait 0 0 1 2.54e3 0.00 + .... + --------------- + Thread 1 vpp_wk_0 (lcore 9) + Time 6152.9, average vectors/node 1.00, last 128 main loops 0.00 per node 0.00 + vector rates in 1.3073e2, out 1.3073e2, drop 6.5009e-4, punt 0.0000e0 + Name State Calls Vectors Suspends Clocks Vectors/Call + TenGigabitEthernet86/0/0-outpu active 804395 804395 0 6.17e2 1.00 + TenGigabitEthernet86/0/0-tx active 804395 804395 0 7.29e2 1.00 + arp-input active 2 2 0 3.82e4 1.00 + dpdk-input polling 24239296364 804398 0 1.59e7 0.00 + error-drop active 4 4 0 4.65e3 1.00 + ethernet-input active 2 2 0 1.08e4 1.00 + interface-output active 1 1 0 3.78e3 1.00 + ip4-glean active 1 1 0 6.98e4 1.00 + ip4-icmp-echo-request active 804394 804394 0 5.02e2 1.00 + ip4-icmp-input active 804394 804394 0 4.63e2 1.00 + ip4-input-no-checksum active 804394 804394 0 8.51e2 1.00 + ip4-load-balance active 804394 804394 0 5.46e2 1.00 + ip4-local active 804394 804394 0 5.79e2 1.00 + ip4-lookup active 804394 804394 0 5.71e2 1.00 + ip4-rewrite active 804393 804393 0 5.69e2 1.00 + ip6-input active 2 2 0 5.72e3 1.00 + ip6-not-enabled active 2 2 0 1.56e4 1.00 + unix-epoll-input polling 835722 0 0 3.03e-3 0.00 |