For the first part of your question, regarding PLS vs. SEM, this introductory article may be helpful: www.stat.umn.edu/~sandy/courses/8801/articles/pls.pdf
PLS is an exploratory technique in contrast to SEM which is often used for confirmatory purposes - confirming hypothesised models as in CFA or SEM. PLS can done with much lower sample size - say with 50 whereas SEM requires 200+. PLS is often used for theory development. Collection of large data set may be unwarranted at this stage. The biggest disadvantage of PLS is that it has no universal fit indices.
Kevin, Raj, and George have already provided very helpful information. Raj and George have also presented the distinction between the PLS path modeling (variance based approach) and covariance based SEM path modeling (the conventional SEM). The PLS based path modeling is also a form of the SEM but it differs from the conventional SEM in terms of the estimation method.
To your question, whether or not the PLS is the substitute of the covariance based SEM modeling, my answer is both YES and NO. I would say yes if your data does not meet some of the requirements such as small sample size, number of observations less than variables etc. In such situation one can prefer PLS based SEM. However, if the data meets the requirement of conducting conventional SEM then PLS based approach may be avoided. One reason is that the validity of the model can not be attested with the PLS base approach. In this sense my answer is "NO".
As far as the goodness of fit of the model is concerned some recent researches have made some progress but still the proposed fit indices provide information about how well a PLS path model can explain different sets of data. They do not provide information about the validity of the model itself. The following link to a recent paper may be helpful in this regard.
Thank you all for your valuable inputs. I have worked with SEm eventually. My problem is I am not getting proper result, in the sense that the model is not fitting well, yet in some of the cases covariances are pretty high, in which case I have to conclude about the theoritical support of the model but not the model itself. This has put me in agreat dilemma. In some instances I am getting negative result ( cannot prove my hypothesis) , seriously concerned about this as far as acceptance of my thesis is concerned. Please guide.
I understand your problem. I even faced similar problem when I was working on my research paper. If ur model fit indices are between 0.80 - 0.90 there is a solution. if they are not, there is no solution via SEM. You can manioulate ur sample littele bit and can try. Also if ur relationship is not significant i have a solution for this. Further my suggestion is - go for PLS modelling since technique is not important but PhD is important.
If u want my help please eloborate the results, i will try to help you.