All,
I've been running some simple tests on Ingo Molnar's "fully HT-aware
scheduler" patch and included the initial results at the end of this
message. [ see: http://lwn.net/Articles/8553/ for patch info ]
Some observations...
- Tests with one or two active tasks show significant improvement with
the HT-aware patch. [ See Ingo's comments about this in his patch
notes. ]
- I was a little surprised to see that having as few as three active
(user) tasks is efficient -- at least in the 'make bzImage' case.
- There was quite a bit of variance from run to run with my internal
tests _without_ the HT-aware scheduler. With the patch it was very
consistent.
Note that the "everypoint internal" test is written in Java and is
memory intensive (lots of search). It also has at least two active
threads. It appears that this test corresponds with the best case
for the HT-aware scheduler and the worst for the existing
implementations. Thus my hunt for HT "fixes"...
- With four active folding@home tasks and the HT-patch, %CPU and %user
stats indicate that the tasks are each receiving over 50% of each
(logical) CPU. To be more clear, CPU0+1 shows a total of ~116% user
and CPU2+3 shows ~109%.
-ASM
----------------------------------------------------------------------------------------------------------------
test kernel: 2.5.33-HT-aware 2.5.33 2.4.8-10bigmem notes
---------------------- ------- --------------- ------------- -------------- ------------------------------
make -j 1 bzImage 195 206 210 make the 2.5.33 kernel
+5.3%
make -j 2 bzImage 117 138,147 127 "
+15.2%
make -j 3 bzImage 110 110 110 "
make -j 4 bzImage 109 109 107 "
make -j 5 bzImage 109 109 109 "
everypoint internal bench 150 168,188,218 154,216,224 Sun 1.4.1RC1 "-server -Xms..."
+10.7%
-----------
%over2.5.33
Supermicro P4DPE-G2 1. times are in real time
E7500 chipset 2. multiple times are listed if there is variance
2 x 2.2GHz Xeon 3. otherwise, best times were selected
Hyperthreading enabled in BIOS
8GB RAM
Allan MacKinnon
allanmac@everypoint.net
----------------------------------------------------------------------------------------------------------------
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sun Sep 15 2002 - 22:00:15 EST