Circuit Cellar-Issue 210- has this question "Which connection(s) to a decoupling capacitor should be shortest?why?" in its "Test your EQ section". I couldn't understand what they meant 'Which connection(s)'. Can somebody explain this?
First let's look at what a bypass capacitor is good for. Let's talk about digital circuits, like discrete logic, or an FPGA or a microcontroller. Contrary to what's been said in comments, in many many cases it is a pretty good assumption that the output drivers are the dominant source of power supply switching noise. Exceptions could be very high-end microprocessors or very large FPGAs.
For traditional logic designs it generally also is reasonable to assume capacitive loads. If your outputs are driving transmission lines, you've got some other issues to worry about, but it won't actually change the results of the analysis much.
Let's say an output driver is switching from low to high. Then the high frequency components of current flow basically as shown in this diagram:
When the output is switching from high to low, the high frequency components of current flow basically like this:
I've put the reference ground down at the end of the line because that's where the receiver will be trying to decide between a '1' and a '0', so that's where the noise margins will matter.
First thing to notice is the bypass capacitor doesn't have much to do with what happens for 1-0 transitions. We'll leave the issue of ground bounce for another question.
Second thing is that for the 0-1 transitions, the current path goes from the capacitor to the driver, and it comes back to the capacitor from the signal's return path. So in this case, what's critical is the path from the capacitor to the driving chip's Vcc pin, and from the ground plane (you are using a ground plane, right?) to the capacitor.
Of course if you're doing analog filters, or power, or something else, you should look for the current loops and do an analysis of your specific situation to know how to optimize your bypassing layout.
For reference, here's the original answer to the question:
If any part of this is not clear, I'd be happy to expand on it as needed.
In an ideal world both of the leads of the decoupling capacitors should be as short as possible and be as close to the IC supply pins as possible.
Now that said, with many logic chips using threshold specifications following the old TTL standard where the noise margins near ground are less than those near the VCC or VDD then it is desirable use extra care to keep down noise on the GND bus. Thus if you have a non-ideal condition and it is necessary that the bypass capacitor has a tradeoff of where one connection has to be longer than the other then it will be appropriate to favor the shorter connection at the GND side of the capacitor.
If a circuit used all CMOS type logic thresholds where the noise margins at GND and VDD are the same then it may be best to try to even out the bypass capacitor connection length on both sides.