Add the utility function to clear page cache

[kingcrimsontianyu]

This PR introduces a utility function to clear page cache in C++ and Python.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[kingcrimsontianyu]

/ok to test 7051979

[madsbk]

Can we also try without sudo? Some like: https://github.com/rapidsai/cudf/blob/6bc515d8219538b104861b48f2b8822a53be841f/cpp/benchmarks/io/cuio_common.cpp#L230

[kingcrimsontianyu]

I think cuDF's implementation still relies on sudo or superuser privilege. The logic is:

1. run the cache clearing command in a shell
2. run the sudo-ed command in a shell
3. if both of them fail, throw an exception

Consequently:

• General case
  • For superuser, 1 succeeds.
  • For non-superuser, 1 fails, 2 succeeds once providing the right password.
• Docker container
  • Without --priviledged argument, both fail regardless of superuser or not.
  • With --priviledged, the same with the general case (except that no password prompt for non-superuser).

So this PR simplifies 1 and 2 to a single sudo command, and also returns boolean to indicate success/failure instead of throwing on failure.

PS: I actually had another implementation earlier that does not use sudo at all: 606b233#diff-e00c8157cb029cd5327cfe9b8b834601b40dcfef0b847bc3d6272e3c7b5c3c1cR216
This method does not have the problem in system() or popen() (they create a separate process and run the shell command, which adds to security vulnerability). But it also lacks the convenience: users need to run the entire process with sudo in order to clear the cache, as opposed to only sudo the cache clearing part in the shell process.

PSS: I also had an experimental snippet that does something like below, with the hope that I can achieve temporary privilege elevation for non-superusers:

perform tasks
{
    // temporarily elevate privilege
    // stash the real uid and real gid
    SuperUserRaiiContext ctx;
    perform tasks
} //  set to the original real uid and real gid
perform tasks

But I scratched this idea too, because it still requires users to run the process with elevated privilege, and as a result the code prior to SuperUserRaiiContext is still privileged.

[kingcrimsontianyu]

@vuule explained that the non-sudo command is used on specially configured machines where unprivileged users have no access to sudo executable but can run sysctl. This complicates things, as the sysctl command does not return non-zero code on insufficient permission failure, unlike the other approach echo 3 | tee /proc/sys/vm/drop_caches which does. So I changed to cuDF's implementation that uses stderr to check failure instead.

[vuule]

when would we want to call this without clearing the dirty pages?

[kingcrimsontianyu]

I'd say these options are for curious minds who wonder the respective effect on page cache 😃

[bdice]

Typically we try to keep __all__ alphabetized. Could you sort these lines?

[kingcrimsontianyu]

Thanks. Sorted.

[madsbk]

Update to say it first try without sudo?

[kingcrimsontianyu]

Thanks. The comments have been updated.