The parallel verison of tensor-tools is a parallel tensor-network library currently customized for performing DMRG (with quantum numbers) in parallel. The parallelization is done via MPI over nodes, using CTF tensors. It is particularly powerful for spin-systems or those without conserved U(1) symmetries.
Not only can it reach bond-dimensions which are inaccessible serially, but (if we extrapolate what a serial code could do on a big-enough machine) it achieves a 10x speedup at no extra cost in node-hours. See arxiv:2007.05540.
The parallel version of tensor-tools was developed by Ryan Levy under the guidance of Bryan Clark with CTF and other help from Edgar Solomonik. It is built on top of the serial version of tensor-tools which was written (circa 2017) by Xiongjie Yu (under the guidance of Clark); the serial code now also has additional contributions/bug-fixes from Ryan Levy.
Requirements:
- CTF
- compiled with lapack/scalapack support; hptt support preferred
- example configure line:
./configure --no-dynamic --build-hptt --with-scalapack' 'LIB_PATH=-L${MKLROOT}/lib/intel64/' 'LIBS=-lmkl_scalapack_lp64 -Wl,--start-group -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lmkl_blacs_intelmpi_lp64 -Wl,--end-group -lpthread -lm' - If using sparse tensors, intel MKL is strongly suggested
- HDF5
- Eigen3
- C++11 support (GCC 6.3.0/8.2.0, Intel 18.0.2/18.0.3.222 compilers tested)
After installing the required libraries, simply make a copy of Makefile.in and adjust the variables to their pointed location. Examples of a linux system, Blue Waters, and Stampede2 are provided. Then build (make -j) which will create an executable run based on the project/main.cpp file.
To execute, launch with MPI using mpirun -n [# of procs] run [inputFile] and follow CTF good practice such as export OMP_NUM_THREADS=1 and CTF_PPN=[# of procs/node].
Conversion to and from ITensor files is fully supported for v2 and v3.1.x (to be determined). See [link repository here].
We utilize both the AutoMPO system from ITensor and some of their lattice code, to have better interoperability between the two. See EXTERNAL.md for more.
Some of the general tensor-tools interface/API was also inspired by ITensor.