On Thu, 15 Oct 1998 [EMAIL PROTECTED] wrote:
> of the CPU time. Here's some bonnie observations, after loading the
> same stuff onto the array as was there for the pre-0.90 RAID build:
>
> -------Sequential Output-------- ---Sequential Input--
> -Per Char- --Block--- -Rewrite-- -Per Char- --Block---
> RAID5.5 256 3048 43.1 3261 31.4 1852 31.3 5195 54.7 6155 41.0
> TM1280S 100 2989 44.1 2763 26.2 1177 19.4 2294 23.3 2707 13.2
> TM1280S 256 2500 37.0 2640 25.8 1151 18.5 2500 24.3 2710 13.5
> NEWRAID5 1024 3610 50.8 4140 41.1 2489 43.4 7141 80.8 8482 67.5
> Throughput has improved significantly with the new RAID patches, but at
> a serious CPU cost. [...]
well CPU cost scales linearly with bandwith, this is pretty much normal
with block granularity IO. I get similar bandwith with lower CPU cost,
could you enable kernel profiling and post a readprofile output to see
where those CPU cycles are spent? We could sure optimize that if it's
spent at the wrong place.
checksumming cost only matters for writes, and you have a nontrivial CPU
load for Block _reads_ already, so there is something else going on ...
For that type of bandwith i usually get something in the range of 15-20%
CPU utilization, on a PII 350. (on NCR810 and 875 controllers)
-- mingo
