[ovs-discuss] Discussion on the logical rationality of flow-limit

[Ben Pfaff]

On Fri, Apr 30, 2021 at 06:10:43PM +0800, taoyunupt wrote:
> 
> 
> 
> At 2021-04-29 06:39:11, "Ben Pfaff" <blp at ovn.org> wrote:
> >On Wed, Apr 28, 2021 at 08:12:06PM +0800, taoyunupt wrote:
> >> Hi，
> >>      Recently I encountered a TCP connection performance problem, the test tool is Apache benchmark.
> >>      The OVS  in my environment is set for  hardware offload solution.  The "Requests per second" is about 6000/s， it closed to non-offload solution.
> >> 
> >> 
> >>       "flow-lmit"  has a dynamic balance in udpif_revalidator, it will modify by the OVS condition（which is pind to "duration"）.   In the revalidate function, when the number of flows is greater than twice the "flow-limit" , the delete flow operation will be triggered to delete all flows; when the number of flows is greater than the "flow-limit", the aging time will be adjusted to 0.1s, Slowly delete flow.   
> >> 
> >> 
> >>      
> >>      I found that the reason for the poor performance is that when the number of flows in the datapath increases and the processing power of OVS decreases, a large number of flow deletions are generated. 
> >>      As we know, In the hardware offloading scenario, although there are a lot of flows, in fact, apart from the first packet, there is no need to process subsequent packets. 
> >>      In my opinion, the dynamic balance mechanism is very necessary, but we need to increase the value of “duration”, or provide some new switches for some high-performance scenarios, such as hardware offloading.
> >>      Do we still need to restrict the number of flows so strictly? By the way, do you have another solution to resolve this?   
> >
> >It's been a long time since I worked on this, but I recall two reasons
> >for the flow limit.  First, each flow takes up memory.  Second, each
> >flow must be revalidated periodically, meaning that it uses CPU as
> >well.
> >
> >I don't, off-hand, remember the real reasons why the logic for deleting
> >flows works as it does.  It might be in the comments or the commit
> >messages.  But, I suspect, it is because above the flow-limit we want to
> >try to reduce the amount of memory and CPU time dedicated to the cache
> >and, if we arrive at twice the flow limit, we conclude that that try
> >failed and that we must have a large number of very short flows so that
> >caching is not very valuable anyhow.
> >
> >In a hardware offload scenario, we get rid of some costs (the cost of
> >processing and forwarding packets and perhaps the memory cost in the
> >datapath) but we still have the cost of revalidating them.  When there
> >are many flows, we add the extra cost of balancing flows between
> >software and the offload hardware.
> >
> >Because of the remaining cost and the added ones when there is hardware
> >offload, it's not obvious to me that we can stop limiting the number of
> >flows.  I think that experimentation and measurements would be needed.
> >Perhaps this would be an adjustment to the dynamic algorithm, rather
> 
> >than a removal of it.
> 
> 
> I think we can increase the init `flow_limit` in udpif_create,10000 is a small number for current server and OS, and if 'duration' is small ,we should increase faster by a lager number not `flow_limit += 1000;`.
> I have not better idea for this situation. Do you have some suggestion? I am very glad to do this change.

What kind of number are you thinking about?  I'd like to come up with a
rationale for choosing it.  It might be even better to come up with an
algorithm or a heuristic for choosing it.