Re: HEADS UP: ULE off in 5.3

On Mon, 25 Oct 2004, Scott Long wrote:

> A lot of poeple have noted recently that a lot of bug reports are coming
> in with strange symptoms, and only after a lot of detective work does it
> turn out that the submitter is using the ULE scheduler.  We turned off
> the ULE scheduler months ago and tried to make it clear that it has
> problems and is not considered stable.  Since there are likely many
> people out there that might not be aware of what scheduler they are
> using, I've set the ULE source code to error out and not compile at all
> for RELENG_5 and RELENG_5_3.  The time spent debugging ULE-induced
> problems is causing way too much lost time, and we need to focus on
> debugging the supported configurations, not the unsupported ones. 

FYI for everyone else, as I know Scott knows this already, but a number of
the features that gave ULE higher throughput on some SMP boxes are now
also available with SCHED_4BSD, such as the use of IPIs to wake up idle
CPUs more quickly when there's load to shed.  I've seen dramatic
improvements in scheduling, especially on SMP, with 4BSD over the last few
months, and the results have done good things to the performance numbers
for applications such as MySQL.  4BSD doesn't have everything ULE had,
such as per-cpu run queues, but it's still a good step up from what
shipped in 5.2.

Robert N M Watson             FreeBSD Core Team, TrustedBSD Projects
robert_at_fledge.watson.org      Principal Research Scientist, McAfee Research