Evgeniy writes:
Here forking connector module "exits" and can handle next fork() on the
same CPU.
Fine ... but it's not about what the fork_connector does. It's about
getting the accounting data to disk, if I understand correctly.
That is why it is very fast in "fast-path".
I don't care how fast a tool is. I care how fast the job gets done. If
a tool is only doing part of the job, then we can't decide whether to
use that tool just based on how fast that part of the job gets done.
The most expensive part is cn_netlink_send()/netlink_broadcast(),
with CBUS it is deferred to the safe time,
This is "safe time" for the immediate purpose of letting the forking
process continue on its way. But the deferred work of buffering up the
data and writing it to disk still needs to be done, pretty soon. When
sizing a system to see how many users or jobs I can run on it at a time,
I will have to include sufficient cpu, memory and disk i/o to handle
getting this accounting data to disk, right?
2) Using a modified form of what BSD ACCOUNTING does now:
- forking process appends single fork data to in-kernel buffer
It is not as simple.
It takes global locks several times, it access bunch of shared between
CPU data.
It calls ->stat() and ->write() which may sleep.
Hmmm ... good points. The mechanisms in the kernel now (and for the
last 25 years) to write out BSD ACCOUNTING data may not be numa friendly.
Perhaps there should be a per-cpu 512 byte buffer, which can gather up 8
accounting records (64 bytes each) and only call the file system write
once every 8 task exits. Or perhaps a per-node buffer, with a spinlock
to serialize access by the CPUs on that node. Or perhaps per-node
accounting files. Or something like that.
Guillaume, Jay - do we (you ?) need to make classic BSD ACCOUNTING data
collection numa friendly? Based on the various frustrated comments at
the top of kernel/acct.c, this could be a non-trivial effort to get
right. Maybe we need it, but can't afford it.
And perhaps my proposed variable length records for supplementary
accounting, such as <parent pid, child pid> from fork, need to allow
for some way to pad out the rest of a buffer, when the next record
won't fit entirely.
That work is deferred and does not affect in-kernel processes.
The accounting data collection cannot be deferred for long, perhaps
just a few minutes. Not until the data hits the disk can we rest
indefinitely. Unless, that is, I don't understand what problem is
being solved here (quite possible ;).
And why userspace fork connector should write data to the disk?
I NEVER said it should. I am NOT trying to redesign fork_connector.
Good grief ... how many times and ways do I have to say this ;)?
I am asking what is the best tool for accounting data collection,
which, if I understand correctly, does need to write to disk.
