Worse performance with SCHED_ULE?

Michal Mertl mime at traveller.cz
Wed Aug 25 02:07:32 PDT 2004


On Wed, 25 Aug 2004 yangshazhou at hotmail.com wrote:

> Hi all,
>     I've benched SCHED_4BSD and SCHED_ULE in 4-way Xeon box, 5.2.1-release,
> with lmbench and volanomark. Both benchmarks show performance decreasing.
>                 SCHED_4BSD        SCHED_ULE
> lat_proc fork    570                570
> lat_ctx  10       6.1                14.8
> lat_ctx 150        10.45             16.6
> volanomark        7024                4607
>
>     I built sun jdk1.4.2 with libkse as native thread. If with libc_r as
> native thread, the result is incredible 6161 and 10643, showing SCHED_ULE
> brings big boost with libc_r. libc_r is user-level thread lib, we know. Is
> there anything wrong?
>
>     Is ULE scheduler stable in 5.2.1? We experienced a lot of errors when
> running apache benchmarks.

I believe ULE is stable in -CURRENT unless PREEMPTION is defined (which
isn't by default).

There were some changes to it lately which should fix the performance
degradation in comparison to 4BSD in some cases. The author of ULE is
interested in such reports as yours but only with recent ULE (read
-CURRENT or RELENG_5).

-- 
Michal Mertl


More information about the freebsd-smp mailing list