Since the 4 Cheetah-3's per channel (nearly) saturate the bus (by
the leveling off in performance), I decided to re-test the DAC1164P
don't all-h/w on those 4+4 drives since it should be less to keep
up with than the 10+10 from before
all_hw_4+4
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
2047 20878 92.6 51166 40.8 21132 46.8 25705 94.6 49572 70.6 597.1 6.3
sw_over_2_hw_raid0_4s
-------Sequential Output-------- ---Sequential Input-- --Random--
-Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
2047 22103 99.6 55974 44.9 20778 46.2 23958 89.2 42752 61.2 591.2 4.6
So removing the s/w raid0 over 2 4-drive h/w raid0's to an 8-drive
raid0 (same drives, same channels, same locations) made writing
worse (~same for rewrite) and reading better, about a draw for seeks
Machine totally idle in both of the above cases.
So is kernel profiling not even possible in the kernel's i386 branch?
James
--
Miscellaneous Engineer --- IBM Netfinity Performance Development
