cvs commit: src/sys/i386/conf GENERIC

[Robert Watson]

On Sun, 7 Aug 2005, David Xu wrote:

> I have reverted it, I will turn it on after 6.0-R. but as you are saying 
> the panics on SMP, I have fixed it, as least I can not reproduce it on 
> my dual PIII machine. In my test, ULE always outperforms 4BSD, your test 
> may vary though. Interactive performance is good, but not as good as 
> 4BSD, it has to be improved, that's why I want to turn it on as a 
> default.

In my experience, ULE has generally performed slightly less well for MySQL 
related workloads on SMP P4 systems.  The following supersmack select 
benchmark on a dual-Xeon 2.3GHz system (with HTT enabled), using a default 
install of MySQL 4.1.12 and default libpthread shows a slight but reliable 
decrease in transaction throughput when running:

x HIPPY-SMP-4BSD
+ HIPPY-SMP-ULE
+--------------------------------------------------------------------------+
|          +        +                                         x            |
|+ x    +  ++       +       *   * + ++            x x    xx x x     x     x|
|       |___________MA________|__|________________A_______M____________|   |
+--------------------------------------------------------------------------+
     N           Min           Max        Median           Avg        Stddev
x  12       6682.11       6967.94       6902.26       6873.42     80.551236
+  12       6675.62        6818.4       6752.41     6754.8017     49.581735
Difference at 95.0% confidence
         -118.618 +/- 56.6309
         -1.72575% +/- 0.823911%
         (Student's t, pooled s = 66.8837)

This is with 11 worker threads and 1000 transactions per run.  ULE has 
improved in performance relative to 4BSD over the last couple of years -- 
it used to be quite a bit slower in comparison (10%+).  The above 
measurements are taken after 4BSD's "settling time" -- i.e., in the 
context of large numbers of simultaenous transactions.  In the past I've 
observed that measurements taken before 4BSD has settled (i.e., in the 
first 60 seconds) are slower than with ULE during that period.

For comparison, here's a UP-compiled kernel running the same workload on 
the same hardware:

x HIPPY-UP-4BSD
+ HIPPY-UP-ULE
+--------------------------------------------------------------------------+
|x    x x  xx x      *    +   *x +  x+  ++      + +   +       +     x     +|
|  |_________________MA_____|___________|M_A______________|                |
+--------------------------------------------------------------------------+
     N           Min           Max        Median           Avg        Stddev
x  12       4226.99       4319.02       4254.55     4255.0375     25.105077
+  12        4254.6       4326.84       4282.17     4284.5442     21.002067
Difference at 95.0% confidence
         29.5067 +/- 19.5968
         0.693453% +/- 0.460554%
         (Student's t, pooled s = 23.1447)

ULE comes in slightly faster.  I've not attempted to identify the SMP and 
UP worker thread count sweet spots in either set of tests.

Robert N M Watson