[ovs-dev] [PATCH ovn] ovn-northd: Fix extremely inefficient usage of lflow hash map.

Anton Ivanov anton.ivanov at cambridgegreys.com
Tue Aug 24 17:18:28 UTC 2021


On 24/08/2021 17:25, Numan Siddique wrote:
> On Tue, Aug 24, 2021 at 7:56 AM Anton Ivanov
> <anton.ivanov at cambridgegreys.com> wrote:
>>
>> On 24/08/2021 12:46, Ilya Maximets wrote:
>>> On 8/24/21 1:18 PM, Anton Ivanov wrote:
>>>> On 24/08/2021 12:05, Ilya Maximets wrote:
>>>>> On 8/24/21 7:36 AM, Anton Ivanov wrote:
>>>>>> On 23/08/2021 22:36, Ilya Maximets wrote:
>>>>>>> On 8/23/21 10:37 PM, Anton Ivanov wrote:
>>>>>>>> On 23/08/2021 21:26, Ilya Maximets wrote:
>>>>>>>>> On 8/23/21 10:20 PM, Anton Ivanov wrote:
>>>>>>>>>> Should not be the case.
>>>>>>>>>>
>>>>>>>>>> The map is pre-sized for the size from the previous iterations.
>>>>>>>>>>
>>>>>>>>>> Line 12861 in my tree which is probably a few commits out of date:
>>>>>>>>>>
>>>>>>>>>>          fast_hmap_size_for(&lflows, max_seen_lflow_size);
>>>>>>>>>>
>>>>>>>>>> And immediately after building the lflows:
>>>>>>>>>>
>>>>>>>>>>          if (hmap_count(&lflows) > max_seen_lflow_size) {
>>>>>>>>>>              max_seen_lflow_size = hmap_count(&lflows);
>>>>>>>>>>          }
>>>>>>>>>>
>>>>>>>>>> So the map SHOULD be sized correctly - to the most recent seen lflow count.
>>>>>>>>> Please, re-read the commit message.  It's a first run of the loop
>>>>>>>>> and the 'max_seen_lflow_size' is default 128 at this point.
>>>>>>>> Ack,
>>>>>>>>
>>>>>>>> Not using auto-resizing in single threaded mode is a bug. Thanks for fixing it.
>>>>>>>>
>>>>>>>>     From that perspective the patch is a straight +1 from me.
>>>>>>>>
>>>>>>>>     From the perspective of the use case stated in the commit message- I am not sure it addresses it.
>>>>>>>>
>>>>>>>> If northd is in single-threaded mode and is tackling a GIGANTIC> database, it may never complete the first iteration before the
>>>>>>>> expiration of the timers and everyone deciding that northd is AWOL.
>>>>>>> Well, how do you suggest to fix that?  Obviously, we can always create
>>>>>>> a database that northd will never be able to process in a reasonable
>>>>>>> amount of time.  And it doesn't matter if it's single- or multi-threaded.
>>>>>>>
>>>>>>> In this case NbDB is only 9MB in size, which is very reasonable, and
>>>>>>> northd on my laptop takes more than 15 minutes to process it (I killed
>>>>>>> it at this point).  With the patch applied it took only 11 seconds.
>>>>>>> So, for me, this patch pretty much fixes the issue.  11 seconds is not
>>>>>>> that bad, e.g. ovn-k8s configures inactivity probes for clients to 180.
>>>>>>> It would be great to reduce, but we're not there yet.
>>>>>>>
>>>>>>>> In that case, if it is multi-threaded from the start, it should probably
>>>>>>>> grab the sizing from the lflow table hash in south db. That would be a
>>>>>>>> good approximation for the first run.
>>>>>>> This will not work for a case where SbDB is empty for any reason while
>>>>>>> NbDB is not.  And there is still a case where northd initially connects
>>>>>>> to semi-empty databases and after few iterations NbDB receives a big
>>>>>>> transaction and generates a big update for northd.
>>>>>> A partial fix is to resize to optimum size the hash after lflow processing.
>>>>> I'm not sure I understand what you mean here, because resizing after
>>>>> lflow processing will not help.  The lflow processing itself is the
>>>>> very slow part that we're trying to make faster here.
>>>> That can be the case only with dpgroups. Otherwise lflows are just a destination to dump data and the bucket sizing is irrelevant because there is never any lookup inside lflows during processing. The lflow map is just used to store data. So if it is suboptimal at the exit of build_lflows() the resize will fix it before the first lookup shortly thereafter.
>>>>
>>>> Are you running it with dpgroups enabled? In that case there are lookups inside lflows during build which happen under a per-bucket lock. So in addition to suboptimal size when searching the contention depends on the number of buckets. If they are too few, the system becomes heavily contended resulting in ridiculous computation sizes.
>>> Oh, I see.  Indeed, without dp-groups there is no lookup during
>>> lflow build.  I missed that.  So, yes, I agree that for a case
>>> without dp-groups, re-sizing after lflow processing should work.
>>> We need that for parallel case.
>>>
>>> Current patch (use hmap_insert() that resizes if needed) helps
>>> for all non-parallel cases.
>> Indeed. It should go in.
>>
> Why can't we have hmap_insert() for both parallel and non parallel configs
> to start with and switch over to hmap_insert_fast() when ovn-northd
> has successfully connected to SB DB and has approximated on the
> more accurate hmap size ?

That is possible, but not for dp_groups.

If it is just the lflow compute, you can use hmap_insert, but that does not actually have any benefit. In fact, you will consume much more CPU than merging into a suboptimal hmap and then resizing it at the end.

For dp_groups, the locking is per hash bucket. If you change the number of buckets (as upon resize) your locks are no longer valid and you end up corrupting the data.

I am running tests on dp_groups and I am starting to think that we should abandon the parallelization of lflow compute altogether for the dp_groups case.

I get at best the same results and sometimes worse results. Looking at the the picture on ovn-central node of ovn-heater the threads never ramp up to more than single digit percents - they are waiting on locks. Compared to that the brute-force lflow compute has threads ramping up to 100% and clear benefit from parallelization.

This is for 36 fake nodes, 1800 ports.

A.

>
> Thanks
> Numan
>
>> I will sort out the other cases to the extent possible.
>
>> Brgds,
>>
>> A.
>>
>>> I'm mostly running dp-groups + non-parallel which is a default
>>> case for ovn-heater/ovn-k8s.
>>>
>>>> For the case of "dpgroups + parallel + first iteration + pre-existing large database" there is no cure short of pre-allocating the hash to maximum size.
>>> Yeah, dp-groups + parallel is a hard case.
>>>
>>>> I am scale testing that as well as resize (for non-dp-groups cases) at present.
>>>>
>>>> Brgds,
>>>>
>>>> A.
>>>>
>>>>>> If the sizing was correct - 99.9% of the case this will be a noop.
>>>>>>
>>>>>> If the sizing was incorrect, it will be resized so that the DP searches and all other ops which were recently added for flow reduction will work optimally.
>>>>>>
>>>>>> This still does not work for lflow compute with dpgroups + parallel upon initial connect and without a SB database to use for size guidance. It will work for all other cases.
>>>>>>
>>>>>> I will send two separate patches to address the cases which can be easily addressed and see what can be done with the dp+parallel upon initial connect to an empty sb database.
>>>>>>
>>>>>> Brgds,
>>>>>>
>>>>>> A
>>>>>>
>>>>>>>> A.
>>>>>>>>
>>>>>>>>>> A.
>>>>>>>>>>
>>>>>>>>>> On 23/08/2021 21:02, Ilya Maximets wrote:
>>>>>>>>>>> 'lflow_map' is never expanded because northd always uses fast
>>>>>>>>>>> insertion.  This leads to the case where we have a hash map
>>>>>>>>>>> with only 128 initial buckets and every ovn_lflow_find() ends
>>>>>>>>>>> up iterating over n_lflows / 128 entries.  It's thousands of
>>>>>>>>>>> logical flows or even more.  For example, it takes forever for
>>>>>>>>>>> ovn-northd to start up with the Northbound Db from the 120 node
>>>>>>>>>>> density-heavy test from ovn-heater, because every lookup is slower
>>>>>>>>>>> than previous one.  I aborted the process after 15 minutes of
>>>>>>>>>>> waiting, because there was no sign that it will converge.  With
>>>>>>>>>>> this change applied the loop completes in only 11 seconds.
>>>>>>>>>>>
>>>>>>>>>>> Hash map will be pre-allocated to the maximum seen number of
>>>>>>>>>>> logical flows on a second iteration, but this doesn't help for
>>>>>>>>>>> the first iteration when northd first time connects to a big
>>>>>>>>>>> Northbound database, which is a common case during failover or
>>>>>>>>>>> cluster upgrade.  And there is an even trickier case where big
>>>>>>>>>>> NbDB transaction that explodes the number of logical flows received
>>>>>>>>>>> on not the first run.
>>>>>>>>>>>
>>>>>>>>>>> We can't expand the hash map in case of parallel build, so this
>>>>>>>>>>> should be fixed separately.
>>>>>>>>>>>
>>>>>>>>>>> CC: Anton Ivanov <anton.ivanov at cambridgegreys.com>
>>>>>>>>>>> Fixes: 74daa0607c7f ("ovn-northd: Introduce parallel lflow build")
>>>>>>>>>>> Signed-off-by: Ilya Maximets <i.maximets at ovn.org>
>>>>>>>>>>> ---
>>>>>>>>>>>        northd/ovn-northd.c | 6 +++++-
>>>>>>>>>>>        1 file changed, 5 insertions(+), 1 deletion(-)
>>>>>>>>>>>
>>>>>>>>>>> diff --git a/northd/ovn-northd.c b/northd/ovn-northd.c
>>>>>>>>>>> index 3d8e21a4f..40cf957c0 100644
>>>>>>>>>>> --- a/northd/ovn-northd.c
>>>>>>>>>>> +++ b/northd/ovn-northd.c
>>>>>>>>>>> @@ -4387,7 +4387,11 @@ do_ovn_lflow_add(struct hmap *lflow_map, struct ovn_datapath *od,
>>>>>>>>>>>                           nullable_xstrdup(ctrl_meter),
>>>>>>>>>>>                           ovn_lflow_hint(stage_hint), where);
>>>>>>>>>>>            hmapx_add(&lflow->od_group, od);
>>>>>>>>>>> -    hmap_insert_fast(lflow_map, &lflow->hmap_node, hash);
>>>>>>>>>>> +    if (!use_parallel_build) {
>>>>>>>>>>> +        hmap_insert(lflow_map, &lflow->hmap_node, hash);
>>>>>>>>>>> +    } else {
>>>>>>>>>>> +        hmap_insert_fast(lflow_map, &lflow->hmap_node, hash);
>>>>>>>>>>> +    }
>>>>>>>>>>>        }
>>>>>>>>>>>          /* Adds a row with the specified contents to the Logical_Flow table. */
>> --
>> Anton R. Ivanov
>> Cambridgegreys Limited. Registered in England. Company Number 10273661
>> https://www.cambridgegreys.com/
>>
>> _______________________________________________
>> dev mailing list
>> dev at openvswitch.org
>> https://mail.openvswitch.org/mailman/listinfo/ovs-dev

-- 
Anton R. Ivanov
Cambridgegreys Limited. Registered in England. Company Number 10273661
https://www.cambridgegreys.com/



More information about the dev mailing list