Hi Ales, Mark,

On 10/15/25 10:23 AM, Ales Musil wrote:
> On Fri, Oct 10, 2025 at 9:58 PM Mark Michelson <[email protected]> wrote:
> 
>> On Fri, Oct 10, 2025 at 3:43 AM Ales Musil <[email protected]> wrote:
>>>
>>>
>>>
>>> On Fri, Oct 10, 2025 at 2:51 AM Mark Michelson <[email protected]>
>> wrote:
>>>>
>>>> On Thu, Oct 9, 2025 at 5:25 PM Mark Michelson <[email protected]>
>> wrote:
>>>>>
>>>>> Hi Ales,
>>>>>
>>>
>>>
>>> Hi Mark,
>>>
>>> thank you for the review, but I'm probably missing something here.
>>>
>>>
>>>>
>>>>> I've been thinking through this change, and I think the idea might
>> have some flaws. Let's consider an arbitrary engine node B that takes
>> engine node A as input. In this scenario, engine node B writes to the SB DB
>> and engine node A does not. With this change, we mark engine node B as
>> SB_WRITE, but we do not do the same with node A.
>>>>>
>>>
>>>
>>> Why would we need to? A state doesn't depend on SB being writable.
>>>
>>>>
>>>>> We perform an engine run where recomputation is not allowed because
>> the SB DB is read only. When engine node A runs, since it does not have
>> SB_WRITE specified, it is able to recompute its data. However, when engine
>> node B runs, it cannot recompute its data.
>>>>>
>>>>> Now, what happens during the next engine run?
>>>>> * Let's say engine node A has no new input data. Engine node A's
>> state will be EN_UNCHANGED. In this case, engine node B sees that engine
>> node A is unchanged, so engine node B does not run. However, B really
>> should run because in the previous engine node run, A recomputed. In this
>> case, B's data is incorrect.
>>>>> * Let's say engine node A has new data and can process it
>> incrementally. Engine node A's state will be EN_UPDATED.
>>>>>     * Engine node B may also try to incrementally handle A's tracked
>> data. However, the only tracked data that B will see is from the current
>> engine run. The data that A recomputed in the previous engine run will be
>> unknown to B. In this case, B's data is incorrect.
>>>>>     * Engine node B may not have an input handler for A. In this
>> case, B will recompute and all will be well.
>>>>> * Let's say engine node A has new data and cannot process it
>> incrementally. In this case, A will recompute, and then B will also likely
>> recompute. The data integrity is fine in this case. However, A has had to
>> recompute twice. Before this change, A would only have recomputed on the
>> second engine run.
>>>>>
>>>>> In 2 of the 4 cases, B ends up with incorrect data. In 1 of the 4
>> cases, the data is correct in engine node B, but A has to recompute twice.
>> And in 1 of the 4 cases, the data is correct in engine node B and we
>> successfully avoided an unnecessary recomputation of A. The only time this
>> change has a positive effect is if the SB_WRITE node has no input handlers
>> for its input nodes, and the input nodes are able to incrementally process
>> their data.
>>>>
>>>> I've realized that in the scenario I outlined, a couple of those
>>>> outcomes can't happen. When the first engine run happens, B will fail
>>>> to recompute, and the engine run will be canceled. In ovn-controller,
>>>> if an engine run is canceled, then we call
>>>> engine_set_force_recompute_immediate(). This means that the next run
>>>> will force all nodes to recompute. This means that the second and
>>>> third outcomes (both based on A being able to incrementally process
>>>> its data) cannot happen. Also, when engine_force_recompute is true,
>>>> the state of input nodes is not considered. This means that if A is
>>>> unchanged, it does not matter. B will still recompute and will
>>>> therefore have all of A's data. So my concerns about data integrity
>>>> are not correct.
>>>
>>>
>>>
>>> So none of the outcomes highlighted cannot happen, if B is cancelled
>>> we will immediately run force recompute loop that will set everything
>>> straight.
>>>
>>>>
>>>>
>>>> However, the fact that the engine forces all nodes to recompute means
>>>> that A will recompute on both engine runs. This means that with the
>>>> patch, A will recompute two times across the two engine runs. Without
>>>> the patch, A only recomputes once. So I still think the patch is
>>>> hurting more than it is helping.
>>>
>>>
>>> That's the case only in the worst scenario. Keep in mind that this
>>> doesn't always happen. So it's like all of the sudden we will do two
>>> recmputes for certain nodes back to back. In fact I see the opposite
>>> happening. Only 3 nodes in ovn-controller are marked as SB_WRITE any
>>> of those nodes might cancel but the rest is free to recompute however
>>> they like. We saw that there has been cancels for nodes do not
>>> require SB write and this was harming performance a good example are
>>> nodes pflow_output and lflow_output. Those are very heavy CPU time
>>> consumers, now they are free to recompute even if SB is read only
>>> during the current run.
>>
>> Thanks for the explanation Ales. I was coming at this change with much
>> more familiarity with the northd incremental engine than the
>> ovn-controller incremental engine. In northd, a lot of nodes are
>> eventual inputs for en_northd or en_lflow, both of which write to the
>> SB DB. So in most cases, even if a node doesn't write to the SB DB,
>> then it eventually will feed into a node that does write to the SB DB
>> [1]. In my head, it felt like the "common" case was for this to
>> happen. However, in ovn-controller, a good majority of the nodes do
>> not write to the SB DB and several of those nodes are never inputs
>> (directly or indirectly) to nodes that write to the SB DB. As you
>> pointed out, pflow_output and lflow_output are good examples of this.
>>
>> The situation I've outlined is still possible in ovn-controller. Many
>> nodes are inputs for en_runtime_data. Letting those nodes recompute
>> when en_runtime_data may eventually be denied the right to do so could
>> be a waste of CPU resources.
>>
>> What if the change were altered a bit: if an engine node is marked as
>> SB_WRITE, then any nodes that are upstream from that node (i.e. any
>> node that is a direct or indirect input to that node) is also marked
>> as SB_WRITE by the incremental engine. Meanwhile, any nodes that are
>> downstream from SB_WRITE nodes are not. This way, nodes like
>> pflow_output and lflow_output can recompute even when the SB database
>> is read-only. But nodes that input into runtime_data, route_exchange,
>> and garp_rarp can avoid unnecessary recomputes.
>>
>> What do you think?
>>
> 
> Your suggestion will unfortunately cause the change to not have any
> benefit, in other words the state would be the same as on main right
> now. For example the runtime_data node has several inputs a lot of
> them are DB nodes e.g. sb_chassis. By marking them also as SB write
> or "canceled if recompte is not allowed" if you will, would basically
> mean it would mostly cancel as without the change. The leaf nodes,
> usually the DB ones are always recomputing. They cannot have any
> handlers. So I'm afraid that this change would then have the only
> benefit of documenting which node can write into SB DB.
> 
> But to check if my theory is correct I did run ovn-heater with a
> baseline from main collecting the engine stats. Below you can see
> the results:
> 
> ---------------------------
> main-250-density-heavy-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 1472 ms
> avg     : 129 ms
> 95%     : 537 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 328312207, Compute: 62608243, Cancel: 2067
> (0.0006295836572412308%)
> 
> ---------------------------
> sb-ro-250-density-heavy-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 1343 ms
> avg     : 130 ms
> 95%     : 577 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 325154128, Compute: 62002527, Cancel: 2138
> (0.0006575343247679758%)
> 
> ---------------------------
> main-250-density-heavy-ipv4-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 119 ms
> avg     : 26 ms
> 95%     : 48 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 77772935, Compute: 4420305, Cancel: 627 (0.0008061930541775233%)
> 
> 
> ---------------------------
> sb-ro-250-density-heavy-ipv4-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 131 ms
> avg     : 26 ms
> 95%     : 47 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 77918353, Compute: 4434319, Cancel: 533 (0.0006840493663925366%)
> 
> ---------------------------
> main-250-density-heavy-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 1330 ms
> avg     : 134 ms
> 95%     : 611 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 325845839, Compute: 62012707, Cancel: 1724
> (0.0005290845527722083%)
> 
> ---------------------------
> sb-ro-250-density-heavy-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 1376 ms
> avg     : 133 ms
> 95%     : 569 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 324545521, Compute: 61835316, Cancel: 2680
> (0.0008257701390369827%)
> 
> ---------------------------
> main-250-density-heavy-ipv6-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 174 ms
> avg     : 26 ms
> 95%     : 48 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 76480809, Compute: 4401000, Cancel: 686 (0.0008969570392488919%)
> 
> ---------------------------
> sb-ro-250-density-heavy-ipv6-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 172 ms
> avg     : 26 ms
> 95%     : 48 ms
> total   : 13750
> failures: 0
> 
> IP stats:
> Recompute: 78371328, Compute: 4493119, Cancel: 615 (0.0007847257609313447%)
> 
> ---------------------------
> main-250-density-light-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 1991 ms
> avg     : 79 ms
> 95%     : 124 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 1787284835, Compute: 339628410, Cancel: 3089
> (0.0001728319929486785%)
> 
> ---------------------------
> sb-ro-250-density-light-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 2465 ms
> avg     : 82 ms
> 95%     : 135 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 1744830863, Compute: 334913315, Cancel: 2439
> (0.00013978432246472707%)
> 
> ---------------------------
> main-250-density-light-ipv4-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 123 ms
> avg     : 24 ms
> 95%     : 34 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 47999536, Compute: 5072743, Cancel: 630 (0.001312512687622647%)
> 
> ---------------------------
> sb-ro-250-density-light-ipv4-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 129 ms
> avg     : 24 ms
> 95%     : 34 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 48054016, Compute: 5090679, Cancel: 522 (0.0010862775756348855%)
> 
> ---------------------------
> main-250-density-light-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 2181 ms
> avg     : 79 ms
> 95%     : 141 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 1778682312, Compute: 337995667, Cancel: 1309
> (0.00007359380543499778%)
> 
> ---------------------------
> sb-ro-250-density-light-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 1993 ms
> avg     : 76 ms
> 95%     : 128 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 1775718929, Compute: 337788420, Cancel: 1527
> (0.0000859933390956154%)
> 
> ---------------------------
> main-250-density-light-ipv6-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 179 ms
> avg     : 27 ms
> 95%     : 41 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 47954143, Compute: 4608282, Cancel: 728 (0.0015181170060739068%)
> 
> 
> ---------------------------
> sb-ro-250-density-light-ipv6-ic
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 169 ms
> avg     : 27 ms
> 95%     : 41 ms
> total   : 63750
> failures: 0
> 
> IP stats:
> Recompute: 46764060, Compute: 4505181, Cancel: 597 (0.0012766214054126181%)
> 
> ---------------------------
> main-500-density-heavy-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 3291 ms
> avg     : 483 ms
> 95%     : 1570 ms
> total   : 27500
> failures: 0
> 
> IP stats:
> Recompute: 845825451, Compute: 166132940, Cancel: 2524
> (0.00029840672174334936%)
> 
> ---------------------------
> sb-ro-500-density-heavy-ipv4
> ---------------------------
> ovn-installed:
> min     : 7 ms
> max     : 3223 ms
> avg     : 447 ms
> 95%     : 1504 ms
> total   : 27500
> failures: 0
> 
> IP stats:
> Recompute: 847419493, Compute: 166570179, Cancel: 2072
> (0.0002445070023896653%)
> 
> ---------------------------
> main-500-density-heavy-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 11875 ms
> avg     : 784 ms
> 95%     : 2267 ms
> total   : 27500
> failures: 0
> 
> IP stats:
> Recompute: 857064532, Compute: 167714262, Cancel: 2521
> (0.0002941435453077412%)
> 
> 
> ---------------------------
> sb-ro-500-density-heavy-ipv6
> ---------------------------
> ovn-installed:
> min     : 7 ms
> max     : 7056 ms
> avg     : 598 ms
> 95%     : 2147 ms
> total   : 27500
> failures: 0
> 
> IP stats:
> Recompute: 865415057, Compute: 169182676, Cancel: 2166
> (0.0002502845290800158%)
> 
> ---------------------------
> main-500-density-light-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 6962 ms
> avg     : 306 ms
> 95%     : 1389 ms
> total   : 127500
> failures: 0
> 
> IP stats:
> Recompute: 5688174577, Compute: 1186841715, Cancel: 1468
> (0.00002580792801148937%)
> 
> ---------------------------
> sb-ro-500-density-light-ipv4
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 5508 ms
> avg     : 301 ms
> 95%     : 1325 ms
> total   : 127500
> failures: 0
> 
> IP stats:
> Recompute: 5676845569, Compute: 1184598070, Cancel: 2659
> (0.00004683939289312734%)
> 
> ---------------------------
> main-500-density-light-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 11675 ms
> avg     : 276 ms
> 95%     : 797 ms
> total   : 127500
> failures: 0
> 
> IP stats:
> Recompute: 5591661452, Compute: 1162323896, Cancel: 2665
> (0.00004766025309073021%)
> 
> ---------------------------
> sb-ro-500-density-light-ipv6
> ---------------------------
> ovn-installed:
> min     : 6 ms
> max     : 6673 ms
> avg     : 242 ms
> 95%     : 688 ms
> total   : 127500
> failures: 0
> 
> IP stats:
> Recompute: 5779381065, Compute: 1201598403, Cancel: 1728
> (0.0000298993954640712%)
> 
> The numbers caught me a little by surprise, but the important thing
> is that the percentage of cancels is pretty small in general. So having
> 2 recomputes the input nodes for SB write nodes won't cause any
> significant issue. Also on the other side there doesn't seem to be any
> measurable improvement. However, unless you see a different reason
> I would still like to get this patch in for the documentation purpose.
> 
> 

Thanks, Ales, for testing this!  I agree, even if we don't get a
significant improvement on the ovn-heater benchmarks it's probably still
good to get this in for the additional semantic check it implements.

From my perspective:

Acked-by: Dumitru Ceara <[email protected]>

But I'd love to hear from Mark too before this is merged.

Thanks,
Dumitru

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

Reply via email to