On 2/29/24 01:56, Chris Mi wrote: > On 2/28/2024 8:51 PM, Ilya Maximets wrote: >> On 2/28/24 13:26, Chris Mi wrote: >>> Hi Ilya, >>> >>> On 2/28/2024 7:56 PM, Ilya Maximets wrote: >>>> On 2/28/24 11:03, Eelco Chaudron wrote: >>>>> >>>>> >>>>> On 28 Feb 2024, at 8:06, Chris Mi wrote: >>>>> >>>>>> Hi Eelco, >>>>>> >>>>>> I did a rebase and pushed to github. Nothing changed, but the actions >>>>>> always reported the following failure: >>>>>> >>>>>> -1000 packets transmitted, 1000 received, 0% packet loss, time 0ms >>>>>> +744 packets transmitted, 744 received, 0% packet loss, time 0ms >>>>>> >>>>>> I can't reproduce the failure locally. Not sure if the test machine >>>>>> is a little weak. So how about change the test like this to make test >>>>>> robost: >>>>>> >>>>>> -NS_CHECK_EXEC([at_ns0], [ping -q -c 1000 -i 0.01 -w 12 10.1.1.2 | >>>>>> FORMAT_PING], [0], [dnl >>>>>> -1000 packets transmitted, 1000 received, 0% packet loss, time 0ms >>>>>> +NS_CHECK_EXEC([at_ns0], [ping -q -c 200 -i 0.05 -w 12 10.1.1.2 | >>>>>> FORMAT_PING], [0], [dnl >>>>>> +200 packets transmitted, 200 received, 0% packet loss, time 0ms >>>>> >>>>> It looks like the machine might be too slow to handle the 1000 pings in >>>>> 12 seconds. Scaling it down should not be a problem. Maybe back to 500, >>>>> and see if this is enough? >>> 500 is also not stable. >>>> >>>> Is it necessary to send hundreds of packets? Can we just send 3-10 like >>>> other tests do? >>> 200 seems stable. >> >> But why 5 or 10 is not enough? > If ratio is 1, 5 or 10 is enough. I can change it. > If ratio is 2, we need send more packets. But maybe we only need > to check the range of the sampled packets.
For the ratio of 2 we could send 100 packets and check that the number is within 30-70 range. This gives us less than 0.01% failure rate. With 200 we'll need to check the 70-130 range, which isn't a significantly narrower range and a chance to successfully send 100 pings is much higher than with 200. So, maybe 100 is a good number to test ratio of 2 with? Also, please, use -W instead of -w in pings as -w doesn't stop ping from sending more packets than requested leading to random failures. >> >>>> >>>> I see, there is an existing test that sends 1000 packets, but it probably >>>> should be canged as well. Not as part of this patch set, but in general. >>>> >>>> In case you missed my previous review for this patch where I pointed out >>>> the problem with the slow ping: >>>> >>>> https://patchwork.ozlabs.org/project/openvswitch/patch/20230619050557.310690-9-...@nvidia.com/#3136718 >>> OK, I'll address other comments. >>>> It should be reproducible locally under ASAN.OK. >>>> >>>> Ping is not a reliable tool when we're trying to send a particular number >>>> of packets, especially with intervals under 100ms. >>> OK, I see. >>> >>> Thanks, >>> Chris >>>> >>>> Best regards, Ilya Maximets. >> _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
