On 2/6/20 7:39 AM, Vishal Deep Ajmera wrote:
>>>
>>> So, the root cause of all the issues in this patch, in my understanding,
>>> is the fact that you need to collect statistics for all the bond hashes
>>> in order to be able to rebalance traffic.  This forces you to have access
>>> to PMD local caches.
>>>
>>> The basic idea how to overcome this issue is to not have PMD local bond
>>> cache, but have an RCU-protected data structure instead.
>>>
>>> Memory layout scheme in current patch consists of 3 layers:
>>>   1. Global hash map for all bonds. One for the whole datapath.
>>>      Protected by dp->bond_mutex.
>>>   2. Hash map of all the bonds. One for each PMD thread.
>>>      Protected by pmd->bond_mutex.
>>>   3. Hash map of all the bonds. Local copy that could be used
>>>      lockless, but only by the owning PMD thread.
>>>
>>> Suggested layout #1:
>>>   Single global concurrent hash map (cmap) for all bonds. One for the 
>>> whole
>>>   datapath.  Bond addition/deletion protected by the dp->bond_mutex.
>>>   Reads are lockless since protected by RCU.  Statistics updates must be
>>>   fully atomic (i.e. atomic_add_relaxed()).
>>>
>>> Suggested layout #2:
>>>   One cmap for each PMD thread (no global one).  Bond addition/deletion
>>>   protected by the pmd->bond_mutex.  Reads are lockless since protected
>>>   by RCU.  Statistics updates should be atomic in terms of reads and 
>>> writes.
>>>   (non_atomic_ullong_add() function could be used).
>>>   (This is similar to how we store per-PMD flow tables.)
>>>
>>> #1 will consume a lot less memory, but could scale worse in case of too 
>>> many
>>> threads trying to send traffic to the same bond port.  #2 might be a bit
>>> faster and more scalable in terms of performance, but less efficient in
>>> memory consumption and might be slower in terms of response to slave
>> updates
>>> since we have to update hash maps on all the threads.
>>> Both solutions doesn't require any reconfiguration of running PMD threads.
>>>
>>> Note that to update cmap entry, you will likely need to prepare the new 
>>> cmap
>>> node and use a cmap_replace().  Statistics copy in this case might be a 
>>> bit
>>> tricky because you may lost part of additions within the period while PMD
>>> threads are still using the old entry. To avoid that, statistics copying
>>> should be RCU-postponed.  However, I'm not sure if we need highly accurate
>>> stats there.
>>>
> 
> Hi Ilya,
> 
> I am inclined to use option 2 to have per-pmd cmap of bonds however,
> I see that even in this option we will need to have one cmap (or hmap)
> of bonds at global datapath level. This will take care of 
> reconfigure_pmd_threads
> scenario. New PMDs needs to be configured with bond cmap from global map.

OK. Make sense for a newly created threads.


_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to