To my understanding, practically it should be after the resampler. Lets say, the interpolation factor is 4 at the transmitter side, that means 4 samples per symbol. So , if Symbol rate is 2 MHz, the sample rate is 8 MHz.
Now, in the receiver side,lets say the ADC I am using is sampling at 40MSps i.e the sample rate is 40 MHz, which is 20 times the symbol rate, or 5 times the transmitter sample rate. It is only after resampling that we will get back our original transmitter sampling rate, so that SRRC filter (with oversampling rate 4) can be employed. Otherwise, if we use before the resampler, we have to use a SRRC filter (with oversampling rate 20), which will increase the number of filter coefficients and hence the hardware cost.
Is my understanding correct?
I'm not sure I understand the question. Your transmitter is creating each symbol with 5 data points. Your receiver is detecting each symbol with 20 data points. You don't really care about the transmitter waveform, you care about the symbol. In order to get the symbol, you have to know the timing. In principle it doesn't matter what order you do things in. In practice, noise is a huge problem and the order makes a big difference. In the different articles, did they talk about noise?
You can easily model both methods, add Gaussian noise to your symbol data, and see what works best. It might be the cheaper way works best - but I doubt it :-)
This is one of those six of one, half a dozen of the other decisions that often pop up in communications design. I strongly suspect that in the case of the one that follows the resampler, there is a more generic (possibly easily realized) anti-alias filter in the resampler. Then the matched filter just "cleans up" what comes out of the resampler.
The reason that I say six of one, etc., is because many of this sort of decisions end up using roughly the same amount of physical resources, they just move them around. When that's not the case, and when you see multiple methods in use, it's because each offers its own advantages. In that latter case the system designer needs to decide which collection of advantages and disadvantages is the best fit for the problem at hand.
Or, the author of each text felt that their way was the best way to present the concept, and figured that you'd figure out on your own that there's more than one way to skin a cat.
Theoretically, the matched filter (MF) is used to mitigate the effect of noise. This is done by matching the impulse response of the filter to the transmitted waveform, hence the name. To do that, the timing and frequency must be estimated. So in general, pilot symbols are sent that are used to enable the required estimation. So think of it as having a loop at the receiver consisting of the MF and the timing and frequency estimator. So it depends on where you put the MF in the loop. Usually, the estimators are in the feedback branch.