LR3 "Extra Processing in Develop" Performance Problem

Question

I have been investigating a specific LR3 performance problem.  It may explain a small subset of the problems people have reported in the "Why is LR3 So Slow?" thread.   I'm starting this thread to focus on this particular problem.  I hope others will confirm/refute/refine my findings.
The Problem
In Develop, when I make an adjustment, normally the following happens: The CPU usage (as shown in Activity Monitor's bar graph) jumps to between 50 and 75% for all four cores, the updated image appears, and the CPU usage settles back down.  This all happens in less than half a second.  Note: this is with the image at the Fit size.  However, sometimes I instead get the following after an adjustment: the CPU usage jumps to 50 to 75% for all four cores and the updated image appears as usual, however, instead of settling back down, the CPU usage jumps up to 90 to 100% for all cores and stays there for 3 to 5 seconds before settling down. Thus it appears that LR is doing some kind of "extra processing" since a lot of computation is happening AFTER the updated image has already appeared.  I will refer to this problem as "EP".  Obviously, when you are getting EP, editing in Develop becomes very balky.
Dependency on ratio between image size and displayed size
It appears that EP only happens when the displayed size of the image (in Fit zoom level and perhaps also Fill zoom level) is above a certain percentage of the actual image size (as currently cropped).  Evidence: When editing full 21MP 5D2 images, I don't experience EP.  If I crop the 5D2 image fairly significantly, then I can get EP.  When editing 10MP images from my Canon S90, I usually get EP for landscape orientation pictures but not for portrait orientation pictures (since in Fit mode, landscape images display at a higher zoom level than portrait images).  If I am getting EP, I can eliminate it by sufficiently reducing the size that LR is displaying the image by resizing the LR window smaller, opening additional panels (I normally edit with only the right panel open), displaying the toolbar, etc.  It appears that EP is enabled when the displayed image is about 50% or larger w.r.t. the actual image (as currently cropped).  For example, EP becomes enabled when a 3648 pixel wide S90 image is displayed at least 17 and 7/8 inches wide on my 100 ppi monitor (i.e. about 1787 pixels).
Dependency on HOW an adjustment is invoked
Even when the displayed image size is large enough w.r.t. the actual image size to enable EP, whether you get it on a given adjustment depends on how you invoke it:
- If you CLICK (i.e. press the mouse button down and quickly release it) on the track of one of the sliders (a technique I use often to make big jumps), EP will happen.
- If you press the mouse button down on a slider handle, drag it to a new position, and quickly release the mouse button), EP will happen
- If you press the mouse button down on a slider handle, drag it to a new position, but continue to hold the mouse button down until the displayed image is updated, EP does NOT happen (either before or after you then release the mouse button).
- If you highlight the numeric field at the end of a slider and use the arrow keys (possibly along with Shift) to increment or decrement the value, EP does NOT happen.
- EP will happen if you resize the LR window such that the displayed image size is above the threshold.  (In fact, I determined the threshold by making a series of window width increases until I saw EP indicated by the CPU bar graphs.)
- EP can happen with local adjustment brush applications, but as with the sliders, it depends on HOW you perform the brush stroke.  Single click and drags with immediate mouse release cause EP, drags with delayed mouse button release don't.
- Clicking an earlier History state causes EP
- More exploration could be done.  For example, I haven't looked at Graduated Filter and Spot Removal adjustments.
My theory of what's happening
With LR2, my understanding is that in Develop mode when the displayed image is below 1:1 zoom level, after an adjustment is invoked, LR calculates the new version of the image to display using a fast, simplified algorithm that doesn't include the more computationally intensive adjustments like Sharpening and Noise Reduction (and perhaps works on a lower rez version of the image with multiple sensels binned together?).  It appears that in conditions described above, LR3 calculates the initial, fast image update and then goes on to do the full update of the image, including the computationally intensive adjustments.  Evidence:   setting Sharpening Amount and Luminance and Color Noise Reduction to zero eliminates EP (or reduces the amount of time it takes to be barely noticeable).  I'm not sure whether the displayed image is updated with the results of the extra processing.  I think the answer is Yes since when I tried an adjustment of changing Sharpening Amount from 0 to 90, the initial update of the displayed image showed sharpening but after the EP, the displayed image was updated again to show somewhat different sharpening. Perhaps Adobe felt that it would be useful to see the more accurate version of the image when it is at or above 50% zoom.  Maybe the UI is supposed to cancel the EP if you start to make another adjustment before it has completed but the canceling doesn't happen unless you invoke the adjustment in one of the ways described above that doesn't cause EP.  
Misc
- EP doesn't seem to happen for Process 2003
- As others have mentioned, I'm surprised that LR (both version 2 and 3) in 64bit mode doesn't use more available RAM.  I don't think I've seen LR go above 4GB of virtual memory or above 3GB "Real Memory" (as reported by Activity Monitor) even though I have several GB free.
- It should be obvious from the above that if you experience EP, there are workarounds: reduce the size of the displayed image (e.g. by window resizing), invoke adjustments in ways that don't cause EP, turn off Sharpening and Noise Reduction until the end of editing an image.
System specs
First generation Intel Mac Pro with two dual-core CPUs at 2.66 Ghz
OS 10.5.8
21GB RAM
ACR cache on volume striped across 3 internal SATA drives
LR catalog and RAWs on an internal SATA drive
30" HP LP3065 monitor (2560 pixels wide)
NVIDIA GeForce 7300 GT

Answer 1

I'm impressed by your thorough analysis.
Clearly, the programmers haven't figured out the best way to do intelligent caching and/or parallel rendering at a reduced size yet.
In my experience reducing the settings in the "Details" panel doesn't help.
What really bugs me is that the lag (or increasing lack of interactiveness) depends on the number of adjustments one has made.
This shouldn't be the case. If a cache is produced then every further adjustment should only cost the effort for that latest adjustment and not include adjustments before it. There are things that stand in the way of straightforward edit applications:
If you work below 1:1 preview, adjustments have to be shown in a reduced form. If you don't have a way to faithfully mimic the adjustments on the reduced size, you have to do them on the original image and then scale down. That's expensive.
To the best of my knowledge LR uses a fixed image pipeline. Hence, independently of the order in which you apply edits, they are always performed in the same fixed order. Say all spot removal operations are done first. If you have a lot of adjustment brush edits and then add a spot removal operation, it means that all the adjustment brush operations have to be replayed each time you do a little adjustment on your spot removal edit.
I believe what you are seeing is mostly related to 1.
I also believe that the way LR currently handles a moderate number of edits is unacceptable and incompatible with the notion that it is usable in a commercial setting for more than trivial edits. I suspect there is something else going on. If everyone saw the deterioration in performance after a number of edits that I see, I don't think LR would be as accepted as it is. Having said that, I've read that the problem of repeated applications of the adjustment brush slowing LR down has existed for a long time. I truly hope that this doesn't mean we'll have to live for it for the foreseeable future.
There are two ways I can see how 2. should be addressed:
combine the effects of a set of operations into one bitmap operation. Instead of replaying all adjustment brush strokes one after the other (speedwise it feels like this is happening), compute a single bitmap operation that combines all effects.
give up the idea that there is an image pipeline with a fixed execution order.
Some might argue that the second point is at odds with the whole idea of parametric editing, but I dispute that. Either edit operations are commutable in which case the order is immaterial, or they are not. If they are not, the user applies the edits in a way as he/she sees fit and will thus compensate for any effect of a changed ordering.
N.B., currently the doctrine of "fixed ordering of edit applications" results in the effect that even if you convert an image into B&W all your adjustment brush edits that applied colour tints will still show through. Reasoning: The user should be able to locally tint a B&W image. I agree with the latter but this could be achieved by only applying those tinting brush strokes that were created after the B&W conversion. All the ones that happened before should only be used to obtain the correct luminance values for the B&W conversion but obviously they shouldn't cause tinted areas.
The above example demonstrates to me that users naturally expect operations to occurr in the order they have been introduced, not in a fixed predefined order. If that principle were followed, I see no reason why the speed of a single edit should depend on the number of edits that were done to the image before.
I hope the programmers can (and the management wants to) address the performance issues. While I find LR usable for pretty modest edits, in no way the performance on my system approaches that would I would expect from an industrial strenght application.
P.S.: Your message reminded me of the following: When I experience serious lag with LR showing the strokes I make with an adjustment brush, it helps to pause a moment after the first click before one starts moving. This allows LR to catch up and then one can see the effect of the application pretty much interactively. Otherwise, there is terrible lag and the feedback where you have brushed an effect comes way too late.