Par vs. Seq decomposition experiment update

[Aleksey Shipilev]

On 12/10/2012 08:21 PM, Brian Goetz wrote:
> Has the N*Q isoline moved closer to the origin compared to previous
> experiments?

Short version: you might argue the answer is "yes".

Long version: previous experiments were using "mean" as the primary
metric, which involved GCs in the metric. If we compare the means for
old experiment and the mean for new experiment, then we have up to ~10%
improvement (e.g. 440us vs 400us now), but that could be easily hijacked
by rogue GCs. We do know the new FJP answer for ~10-15% improvements in
cold start scenarios, so this seems to be related.

In any case, there seem to be nothing pathological with FJP at this
point, and I should really focus to the internal lambda mechanics now.
The benefits for isoline should be more fruitful there.

> Separately, I have a good understanding of what N=10^3 means.  But how
> much "real computation" does Q=10^2 mean?

You can easily infer that. With N=10^3 and Q=10^2, seq runs the job at
450us. Assuming the absence of pipelining effects (very crude), we can
figure out N=1, Q=1 is closer to 4.5ns. From the code perspective, the
elementary operation corresponding to Q=1 is just a simple
multiply-add-mask, and it takes around 10 cycles per op.

-Aleksey.