Short Description: A Family of Lightweight TLS implementations
- Remember to set the path to the TLS library: $ export LIB_TLS_ROOT=full_path_to_LibTLS
############################### ########### REMARKS ########### ###############################
-
Folder "LibTLS" contains the PolyLibTLS library. Folder "Demo" contains a (simple) motivating example. Folder "TestIDEA" contains the IDEA benchmark. Folder "LCPCbench" contains several benchmarks that were parallelize by hand under TLS (IDEA, NeuralNet, FFT, BH, EM3D, TSP). Each bench has a very simple "Makefile" and further (short) explanation is given in "RunHint".
-
The library was tested mainly on simple examples that did not require separate compilation of 'cpp' files.
-
For good performance most functions need to be inlined -- i.e.
specLD' andspecST'. -
In some cases the serial-commit model performs poorly, mainly because: a) it uses an `mfence' instruction to guarantee sequential consistency; this can be eliminated in the same way as with the In-Place model.
b) in the presence of iteration-independent RAW dependencies: -- we are currently betting that there are no iteration-independent, hence we provide some light-weight ways to test that: (i) maintain the highest written address and bet that the next read addresses is higher (ii) similar for the lowest (iii) keep a vector maintaining the written addresses; if a vector location is not set then that address has not been written and its value is returned from memory.
In the cases when iteration-independent RAW dependencies are often we would need (to implement) a new Serial-Commit model which maintains its updates in a hash structure (instead of a buffer) and hence can provide the requested value in O(1) time.
-
To run speculation inside a loop you need to do smth like:
for(int j=0; j<4; j++) { ttmm.speculate (); ttmm.cleanUpAll(); }