build(jemalloc): enable the profiling option when compiling the jemalloc

[git-hulk]

This closes #3000.

This only enables profiling at the compile stage, but since we set prof_active to false, the profiling is not active, so it won't incur any overhead for now. We can enable the profiling by setting the environment MALLOC_CONF like:

MALLOC_CONF=prof:true,prof_active:true,lg_prof_interval:30,lg_prof_sample:21,prof_prefix:/tmp/jeprof ./kvrocks

And it will produce the jemalloc under the /tmp directory like the following:

jeprof.2359.0.i0.heap
jeprof.2359.1.i1.heap  
jeprof.2359.3.i3.heap

[Copilot]

Pull Request Overview

This PR enables the jemalloc profiling option at compile time while keeping profiling inactive by default, allowing future activation without performance overhead.

• Introduces a CMake flag that conditionally adds the profiling option.
• Adjusts the configure command to include the profiling flag when profiling is enabled.

[PragmaTwice]

this is a CMake variable and will not be passed to the build phase of jemalloc.

[git-hulk]

Oh? I see Jemalloc's configure.ac recognizes this environment, so I would like to test if CMake could pass this.
https://github.com/jemalloc/jemalloc/blob/edaab8b3ad752a845019985062689551cd6315c1/configure.ac#L1278

[PragmaTwice]

Seems it can also be configured in runtime via MALLOC_CONF env var, so we can ignore it?

[git-hulk]

This line mainly uses to set prof_active to false by default, so that it won’t collect the profiling unless user sets the MALLOC_CONF to enable it.

[git-hulk]

I have tested on my side, the heap profiling won't run if we didn't enable opt.prof=true via MALLOC_CONF, so we can remove this line.

[sonarqubecloud]

Quality Gate failed

Failed conditions
40.6% Coverage on New Code (required ≥ 50%)

See analysis details on SonarQube Cloud

[PragmaTwice]

Is there any performance impact of this flag? e.g. through redis-benchmark.

[git-hulk]

Is there any performance impact of this flag? e.g. through redis-benchmark.

In summary, there's no impact on performance. Many other projects are also enabled when compiling, like ClickHouse and MySQL(Percona)，

The first one is running without enable the --enable-prof when compiling.

hulk@instance-20241021-041032:~$ redis-benchmark -c -p 6666 -n 1000000 -r 1000000 -t set -d 128
====== 6666 -n 1000000 -r 1000000 -t set -d 128 ======
  100000 requests completed in 4.52 seconds
  0 parallel clients
  3 bytes payload
  keep alive: 1
  host configuration "save": 900 1 300 10 60 10000
  host configuration "appendonly": no
  multi-thread: no

99.62% <= 0.1 milliseconds
99.95% <= 0.2 milliseconds
100.00% <= 0.3 milliseconds
100.00% <= 0.7 milliseconds
100.00% <= 1.0 milliseconds
22114.11 requests per second

The second one is running with --enable-prof, but prof_active is false.

hulk@instance-20241021-041032:~$ redis-benchmark -c -p 6666 -n 1000000 -r 1000000 -t set -d 128
====== 6666 -n 1000000 -r 1000000 -t set -d 128 ======
  100000 requests completed in 4.52 seconds
  0 parallel clients
  3 bytes payload
  keep alive: 1
  host configuration "save": 900 1 300 10 60 10000
  host configuration "appendonly": no
  multi-thread: no

99.62% <= 0.1 milliseconds
99.95% <= 0.2 milliseconds
99.99% <= 0.3 milliseconds
100.00% <= 0.4 milliseconds
100.00% <= 0.5 milliseconds
100.00% <= 0.6 milliseconds
100.00% <= 1.0 milliseconds
22128.79 requests per second

The last one is running with --enable-prof and prof_active is true.

hulk@instance-20241021-041032:~$ redis-benchmark -c -p 6666 -n 1000000 -r 1000000 -t set -d 128
====== 6666 -n 1000000 -r 1000000 -t set -d 128 ======
  100000 requests completed in 4.54 seconds
  0 parallel clients
  3 bytes payload
  keep alive: 1
  host configuration "save": 900 1 300 10 60 10000
  host configuration "appendonly": no
  multi-thread: no

99.51% <= 0.1 milliseconds
99.94% <= 0.2 milliseconds
99.99% <= 0.3 milliseconds
100.00% <= 0.4 milliseconds
100.00% <= 0.4 milliseconds
22011.88 requests per second