I got the following benchmarks for AbinitioRelax and relax (version 3.2). The speed improvement I observed compared to 3.1 are so important that I am wondering if they are real. I would appreciate if someone could confirm that such improvement (3.2 vs 3.1) is really to be expected.
System : 90 residues protein
Protocols tested :
1. "AbinitioRelax -abinitio -out:nstruct 100"
2. "relax -relax" (3.1) or "relax -relax:thorough" (3.2) on above abinitio structures
3. "relax -relax:fast" on above abinitio structures
Benchmarks were all done on a single node (dual Intel Nahalem-EP at 2.8 GHz, 24 GB RAM, ), with a single process per node (i.e. yes, I am wasting 7 cores... until I get the pseudo-mpi version of Rosetta compiled).
Linux r107-n37 2.6.18-194.32.1.el5 #1 SMP Wed Jan 5 17:52:25 EST 2011 x86_64 x86_64 x86_64 GNU/Linux
All compiles were static, gcc=4.4.2, and included the following CCFLAGS
CCFLAGS = -pipe -ffor-scope -W -Wall -pedantic -Wno-long-long --param inline-unit-growth=1000 --param large-function-growth=50000 -O3 -ffast-math -funroll-loops -finline-functions -finline-limit=1133 -s -Wno-unused-variable
The following computation times are average user cpu time in seconds/structure for the above three protocols.
Rosetta 3.1 : 26.2 - 84.0 - 30.2
Rosetta 3.2 : 19.8 - 28.2 - 10.2
-abinitio (centroid): 24% faster with Rosetta-3.2
-relax:thorough (3.2) vs -relax (3.1) : 66% faster with Rosetta-3.2
-relax:fast : 66% faster with Rosetta-3.2
I realize from the docs that the relax protocols have changed significantly in 3.2, but 66% faster is amazing.
Also, does anyone know if -relax (default in 3.1 is "slow") is the same as -relax:thorough (in 3.2)?
Even the generation of low-res decoys (AbinitioRelax -abinitio) is quite faster in 3.2... is that normal?
Finally, it appears now that the full-atom refinement (step 2) is must faster than the initial generation of decoys (10.2 sec vs 19.8 sec per structure). Is that real?
Although the results look fine, I would appreciate if someone would confirm that the full-atom refinement is not so compute-intensive in Rosetta-3.2. If so, this is really great!