Re: [PULL 3/9] tests/migration: Reduce autoconverge initial bandwidth

[Michael S. Tsirkin]

On Wed, Mar 25, 2020 at 01:16:26PM +0000, Dr. David Alan Gilbert (git) wrote:
> From: Philippe Mathieu-Daudé <philmd@redhat.com>
> 
> When using max-bandwidth=~100Mb/s, this test fails on Travis-CI
> s390x when configured with --disable-tcg:
> 
>   $ make check-qtest
>     TEST    check-qtest-s390x: tests/qtest/boot-serial-test
>   qemu-system-s390x: -accel tcg: invalid accelerator tcg
>   qemu-system-s390x: falling back to KVM
>     TEST    check-qtest-s390x: tests/qtest/pxe-test
>     TEST    check-qtest-s390x: tests/qtest/test-netfilter
>     TEST    check-qtest-s390x: tests/qtest/test-filter-mirror
>     TEST    check-qtest-s390x: tests/qtest/test-filter-redirector
>     TEST    check-qtest-s390x: tests/qtest/drive_del-test
>     TEST    check-qtest-s390x: tests/qtest/device-plug-test
>     TEST    check-qtest-s390x: tests/qtest/virtio-ccw-test
>     TEST    check-qtest-s390x: tests/qtest/cpu-plug-test
>     TEST    check-qtest-s390x: tests/qtest/migration-test
>   **
>   ERROR:tests/qtest/migration-test.c:1229:test_migrate_auto_converge: 
> 'got_stop' should be FALSE
>   ERROR - Bail out! 
> ERROR:tests/qtest/migration-test.c:1229:test_migrate_auto_converge: 
> 'got_stop' should be FALSE
>   make: *** [tests/Makefile.include:633: check-qtest-s390x] Error 1
> 
> Per David Gilbert, "it could just be the writing is slow on s390
> and the migration thread fast; in which case the autocomplete
> wouldn't be needed. Perhaps we just need to reduce the bandwidth
> limit."
> 
> Tuning the threshold by reducing the initial bandwidth makes the
> autoconverge test pass.
> 
> Suggested-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
> Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com>
> Message-Id: <20200323184015.11565-1-philmd@redhat.com>
> Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
> Tested-by: Alex Bennée <alex.bennee@linaro.org>
> Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>


This slows make check down significantly for me, it stays
at the migration test for minutes.

I'm carrying a revert at top of my tree for now but I'd rather
not need that.


This seems like a fragile way to test things anyway.
What happens if someone slows writing even more
e.g. because it's running in a container or a VM?

How about detecting that migration finished too early
and slowing it down until autocomplete triggers?



> ---
>  tests/qtest/migration-test.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/tests/qtest/migration-test.c b/tests/qtest/migration-test.c
> index 3d6cc83b88..2568c9529c 100644
> --- a/tests/qtest/migration-test.c
> +++ b/tests/qtest/migration-test.c
> @@ -1211,7 +1211,7 @@ static void test_migrate_auto_converge(void)
>       * without throttling.
>       */
>      migrate_set_parameter_int(from, "downtime-limit", 1);
> -    migrate_set_parameter_int(from, "max-bandwidth", 100000000); /* ~100Mb/s 
> */
> +    migrate_set_parameter_int(from, "max-bandwidth", 1000000); /* ~1Mb/s */
>  
>      /* To check remaining size after precopy */
>      migrate_set_capability(from, "pause-before-switchover", true);
> -- 
> 2.25.1
> 
>