[Bug runtime/2514] Sequence had xxx drops.

["zanussi at us dot ibm dot com" <sourceware-bugzilla at sourceware dot org>]

------- Additional Comments From zanussi at us dot ibm dot com  2006-04-05 16:19 -------
(In reply to comment #2)
> (In reply to comment #1)
> > How large buffers have you tried?
> you mean the buffer size specified by "stap -s"?
> One of my colleagues(Liang Shanshan) said she tried 1m, 2m, 4m, 8m, 16m buffer
> size, but got almost the same result:
> 
> It seems that option "-s" has no effect on sequence dropping.
> 
> first times:(-s 1)
> Sequence had 5189560 drops.
> counthook average is 8342273
> 
> second times:(-s 2)
> Sequence had 6404573 drops.
> counthook average is 8305578
> 
> ...
> 
> forth times:(-s 8)
> Sequence had 3859367 drops.
> counthook average is 8326021
> ...
> 
> > 
> > Those hard-coded VT100 escape codes in stpd/librelay.c and elsewhere should come
> > out, btw.
> > 

Well, it does look like it's having some effect - when you went to 8Mb, you cut
the drops in half.

If you're logging huge amounts of data, though, you'll eventually reach a point
where stpd can't write the data to disk fast enough to keep up.

One thing that worked well for blktrace was to mmap the output buffers - using
this method, it seems to be able to log data at pretty high rates without
dropping any records and with relatively small buffer sizes.  Another thing that
seemed to work well was to keep a list of unwritten sub-buffers in the daemon
which would act as a 'buffer' against temporary backlogs; a relatively small
buffer size could then be used to handle normal expected traffic, but when data
came in too quickly to be written to disk, the list could temporarily absorb the
backlog.

I've been meaning to try something like this with systemtap; sounds like it
would be a good time to do it - I'll play around with it as soon as I get the
chance...

-- 


http://sourceware.org/bugzilla/show_bug.cgi?id=2514

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