(j3.2006) Did we intend to prohibit this?
Tue Mar 14 02:17:40 EDT 2017
On Tue, 2017-03-14 at 14:09 +0900, Cohen Malcolm wrote:
> Why on earth would you run a large code for a long time on many cores to
> verify what happens in one single statement. It would not take very much
> code, time, or cores, to check what happens here.
Because contrived non-representative toy tests don't reveal the total
contribution of each statement to the total runtime.
> If avoiding the memory allocator is your aim (and MOVE_ALLOC does not avoid
> it, but anyway), you say there is a maximum effect of 3% on your runtime.
> That does not sound worth getting all steamed up about. Especially when a
> few simple experiments might shed light on the topic.
Simple experiments give information about one statement, but not about
the effect of executing it the number of times a real profile does.
I reduced total runtime by 40% by chasing dozens of 3% problems. I
thereby reduced the cost of the required computing platform by $200,000,
more than ten times the labor cost I charged to my account to make the
More information about the J3