[SATLUG] AoE results

[Travis H.]

On Tue, Nov 28, 2006 at 05:39:05AM -0600, K. Spoon wrote:
> generating a bunch of files, though, AoE leaves NFS in the dust.  Both
> methods also have the benefit of using less CPU time than raw local ATA
> writes, though.

Less CPU usage perhaps, but network activity is expensive.

Great paper:

http://research.microsoft.com/~gray/papers/MS_TR_99_100_Rules_of_Thumb_in_Data_Engineering.pdf

Rules of thumb:

"[one] message costs 10k instructions and 10 instructions per byte"
"a disk I/O costs 5k instructions and 0.1 instructions per byte"

Maybe network packets are less expensive than this implies because they
are ethernet but not IP, but I suspect that it won't change two orders
of magnitude.

So I suspect the reason why CPU % is going down is that the amount of time
has increased in the AoE case compared to the directly-attached disk.
It's waiting for data from the network, probably.  The latency can't be
lower because the remote system still has to do just as much as a single
system would do with a directly attached disk, plus both systems have to
do a network transaction.

Also, is DMA and other hdparm stuff tuned for this local disk?
That would significantly reduce CPU time on the system for the
local case.
-- 
"Cryptography is nothing more than a mathematical framework for
discussing various paranoid delusions." -- Don Alvarez
<URL:http://www.subspacefield.org/~travis/> -><-