This is the mail archive of the mailing list for the systemtap project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[Bug kprobes/2062] Return probes does not scale well on SMP box

------- Additional Comments From jkenisto at us dot ibm dot com  2006-07-07 23:32 -------
(In reply to comment #15)
> Created an attachment (id=1147)
 --> (
> testing data Jim's patch
> Hi Jim,
>   I manually applied your patch to, and tested on a 8-way ppc64
> box. I use a multi-thread app which will create 8 threads and each thread will
> call getsid() in a loop. The tests shows that the kretprobes still doesn't
> scale well on SMP box. Here is the results:
> <1> without being probed by stap:
> Total cpus: loops = 40000000, average = 5822 ns
> <2> probed by 'stap -e "probe syscall.getsid {}" -bM'
> Total cpus: loops = 40000000, average = 7688 ns
> <3> probed by 'stap -e "probe syscall.getsid.return {}" -bM'
> Total cpus: loops = 40000000, average = 25277 ns

This is troubling.  A kretprobe is more expensive than a kprobe, but it
shouldn't be that much more.  Given uniprocessor performance ratios, I'd expect
numbers in the range of 8,000-10,000 ns, not 25,000 -- if lock contention
weren't a major factor.

What numbers do you get when you run the syscall.getsid.return test with the
"old" version of kretprobes?

> I ever used oprofile to sample the kretprobe, and the sample data is similar
> with the data in this attachment.

The high number for .__spin_yield() confirms that there's significant lock
contention, and the high number for ._spin_lock_irqsave() (compared to
._spin_lock()) suggests that the contention is on the hash-bucket locks
(kretprobe_table_locks[]) rather than the per-kretprobe lock.  This is a
surprise to me.

My version of stap doesn't appear to create any calls to spin_lock_irqsave() in
the handlers.  Could you please verify that that's true for your version?  Run
stap with -p3 and verify that there are no calls to spin_lock_irqsave()?

I did some experimentation and verified that hashing on current() provides a
reasonably good distribution of hash indexes.  That is, a bunch of tasks
launched together but running concurrently seem to hash to different buckets. 
(If you run them consecutively, they tend to re-use the same task_struct and
therefore the same hash bucket... but if they're not concurrent, there'd be no
contention, right?)

A good thing to try next is to factor SystemTap out of the experiment.  I'll
attach a .c file that's equivalent to "probe syscall.getsid.return {}".


------- You are receiving this mail because: -------
You are the assignee for the bug, or are watching the assignee.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]