[ovs-discuss] Flood test with xen/openvswitch

[Ben Pfaff]

On Fri, Sep 09, 2011 at 07:47:52PM +0200, S?bastien Riccio wrote:
> It's more a cpu usage problem than memory, last test i did:
> 
> hping3  -q -S -L 0 -p 80 -i u50 192.168.50.2
> 
>   PID USER      PR  NI  VIRT  RES  SHR S*%CPU* %MEM    TIME+  COMMAND
>  2211 root      10 -10  7660 3816 1076 R *93*  0.5   1:33.17 ovs-vswitchd
>  1793 root      20   0     0    0    0 S   25  0.0   0:24.04 netback/0
> 
> constant 93-95% cpu usage

That's expected behavior.  When new flows constantly pop up, it takes
CPU time to decide what to do with them, and eventually you run out of
CPU time.  This will be true of any kind of smart software bridge,
including the Linux kernel bridge, but it's more obviously visible
with Open vSwitch because the CPU cost gets credited to a specific
process instead of to just a vague "system time" percentage, and
because the cost of a kernel/user transition is higher than when
everything happens exclusively in the kernel

(That's why I've been paying more attention to the memory usage
report.  ovs-vswitchd shouldn't grow without bound.)