On Sat, Dec 19, 2009 at 8:07 PM, Adam Silberstein <[email protected]>wrote:

> ...
> -My bail out code should keep me from going to the bitter end.


Yes.



>  I tried
> bailing out at different numbers (2, 100).  I'm seeing way better
> throughput.


Yeah.  You want to balance rpc calls and over-fetching.



> CPU still strikes me as high, but maybe I'm underestimating
> what's going on under the covers.


We're probably heavy-weight here.  We're doing a bit of fancy footwork so we
ride the scanner across regions and so we do right thing regards caching.



> I have an 8-core machine.  With 10
> parallel clients doing 100 record scans, usage across each core adds up to
> 60-70%   Does that seem reasonable?
>

I'm not sure.  Can you hang a profiler on it to see where the CPU is being
chewed up?

Result is backed by a raw list of KeyValues.  You have access to it even.
 I'd guess that working w/ the raw list, you'd save CPU but not sure you can
do that via Scanner#next (maybe someone else knows better).

Go easy,
St.Ack

Reply via email to