Skip mysql tests on ppc64 runner

[NattyNarwhal]

For whatever reason, mysql tests are apocalyptically slow on ppc64 (presumably -O0 ASAN makes this much worse).

skipSlow is enabled, which skips some other tests, consistent with the Alpine nightly job. Could reconsider enabling some of these even in skipSlow.

[iluuu1994]

The idea of skipSlow was to make the ASAN on-push job run in a reasonable amount of time. I don't know why Alpine uses this flag, possibly copy & paste. This will make the ASAN on-push job less useful for mysql development.

[iluuu1994]

Looking at the number of changes in mysqli/pdo_mysql, I'd be ok with this. Though I'd still prefer trying without ASAN and with -O2 first. I don't know Gentoo, but this doesn't sound too hard? And if we just tried we'd at least know whether it makes the problems go away. I see no benefit of these flags for us.

[mvorisek]

The mysqli should not be that slow even under heavy instrumentation.

Do you know you can run valgrind like in https://github.com/php/php-src/blob/php-8.5.0beta3/benchmark/benchmark.php#L118 to capture a callgrind profile and then analyse it using tools like KCachegrind?

[NattyNarwhal]

The idea of skipSlow was to make the ASAN on-push job run in a reasonable amount of time. I don't know why Alpine uses this flag, possibly copy & paste. This will make the ASAN on-push job less useful for mysql development.

I can mark the Alpine nightly job as not skipSlow if it helps.

Looking at the number of changes in mysqli/pdo_mysql, I'd be ok with this. Though I'd still prefer trying without ASAN and with -O2 first. I don't know Gentoo, but this doesn't sound too hard? And if we just tried we'd at least know whether it makes the problems go away. I see no benefit of these flags for us.

mysqld itself is -O2 no ASAN, it's just mysqlnd that's eating the full performance penalty of ASAN/-O0. I guess we could force -O2 on that extension?

The mysqli should not be that slow even under heavy instrumentation.

Do you know you can run valgrind like in https://github.com/php/php-src/blob/php-8.5.0beta3/benchmark/benchmark.php#L118 to capture a callgrind profile and then analyse it using tools like KCachegrind?

I agree; I captured a flamegraph a while back in April, using perf rather than callgrind (attached). I'll try doing callgrind later.

perf.svg.gz

[iluuu1994]

mysqld itself is -O2 no ASAN, it's just mysqlnd that's eating the full performance penalty of ASAN/-O0. I guess we could force -O2 on that extension?

Are you referring to mysqlnd the library, or ext/mysqli & ext/pdo_mysql here? IMO, mysqlnd as well as all other libraries should be compiled with -O2 and no ASAN, while everything we ship should be compiled with ASAN. ASAN and UBSAN can catch more things without -O2, so it makes sense to stick with -O0 for our own code. At least if I understood you correctly, you're compiling mysqlnd yourself, which I understand is common on Gentoo? If mysqlnd happens to do a lot of arithmetics, which I could imagine for server communication UBSAN may have a significant impact. But it's just a guess, you've also mentioned that ASAN/UBSAN may not be particularly optimized for ppc64.

[NattyNarwhal]

No, ext/mysqlnd (and thus mysqli/PDO_mysql) should be built against the vendored mysqlnd library one in PHP. I could build it against the system version of the mysqlnd library, but I thought we deprecated that?

[iluuu1994]

Oh, I'm sorry. I have very little knowledge of mysql and it shows. I didn't realize mysqlnd was bundled with php-src. In that case, I don't know what other solution there is. 😕

[NattyNarwhal]

Oh, I'm sorry. I have very little knowledge of mysql and it shows. I didn't realize mysqlnd was bundled with php-src. In that case, I don't know what other solution there is. 😕

I guess I could hack up the build system to drop some flags in ext/mysqlnd...

Do you know you can run valgrind like in https://github.com/php/php-src/blob/php-8.5.0beta3/benchmark/benchmark.php#L118 to capture a callgrind profile and then analyse it using tools like KCachegrind?

FWIW I tried this, and Valgrind doesn't like ASan binaries (ASan refuses to load). Without ASan is still pretty slow, but here:
callgrind.phpmysqlnd.txt