Re: cvs commit: src/sys/kern sched_ule.c (fwd)

On 2004-12-13 17:26:10 Scott Long wrote:

> RELENG_5 is the stable branch.  If quality testing goes into ULE in HEAD
> and it's shown to be as stable as 4BSD then we can consider it for
> RELENG_5 in the future.  Given the incredible problems that we had in
> the scheduler leading up to 5.3, I'm not excited about quickly merging
> these things.

I have FreeBSD 5.3 installed on a dual amd64 colo server of mine and have
been experiencing severe issues with the system and the 4BSD scheduler under
heavy MySQL load.  Originally with 5.3-RELEASE these appeared to be kernel
crashes/deadlocks, but unfortunately I never had a dump device configured 
when I was running 5.3-RELEASE and so I don't have a core file to be examined.

However, I've been checking out the sys/ sources from RELENG_5 fairly frequently
and still experience severe issues with the 4BSD scheduler when the system
is under heavy database load.  Namely, while the kernel appears to remain
running and the system continues to respond to pings, all other network
services cease to function.  New TCP connections are accepted, but the
services don't respond, and existing connections time out.

I have found this does NOT occur when the ULE scheduler is used.  I have
(perhaps foolishly) attempted to copy the minimum necessary files to run the
ULE scheduler from the -CURRENT branch and merge them myself into the 5-STABLE
sources, which I believe are sched_ule.c and kern_switch.c, and have modified
the proc_fini() function in kern_proc.c to panic if invoked (since according to
the comments, UMA should ensure that proc_fini is never called, correct?).  If
these are all the changes that are needed to import the ULE scheduler, then why
continue to include the broken ULE scheduler with an #error tag rather than
importing the minimum sources required for the ULE scheduler to work and leave
it off per default?

I, for one, am experiencing better system stability with ULE than with 4BSD.
If anyone cares to examine my system I can provide shell access.

Tony Arcieri