Employing Multiple CUDA Devices to Accelerate LTL Model Checking

Abstract

Recently, the CUDA technology has been used to accelerate many computation demanding tasks. For example,in [7] we have shown how CUDA technology can be employed to accelerate the process of Linear Temporal Logic (LTL) Model Checking. While the raw computing power of a CUDA enabled device is tremendous, the applicability of the technology is quite often limited to small or middle-sized instances of the problems being solved. This is because the memory that a single device is equipped with, is simply not large enough to cope with large or realistic instances of the problem, which is also the case of our CUDAaware LTL Model Checking solution. In this paper we suggest how to overcome this limitations by employing multiple (two in our case) CUDA devices for acceleration of our fine-grained communication-intensive parallel algorithm for LTL Model Checking.

Designing fast LTL model checking algorithms for many-core GPUs

“…We were able to adapt our implementation of both CUDA accelerated algorithms to mimic this behavior as well. The idea is very similar as in our previous papers [2,8]. In particular, we let the CPU perform (parallel) state space generation while having the GPU apply CUDA accelerated algorithms on partially constructed graph.…”

“…This is slightly tricky due to the space limitation since we cannot afford to create the copy of the whole vector of vertices. We proposed the solution of this problems in [2].…”

“…However, as stated in [2], backward reachability via forward edges is tractable (with certain slowdown), which allows us to implement elimination in the trivial way sketched above. The rest of the algorithm remains unchanged.…”

“…This journal paper provides a survey of the latest advancements in GPU acceleration of LTL model checking including, among others, our own results recently published in several conference and workshop proceedings [8,7,[2][3][4] and exhaustive experimental evaluation. The paper recapitulates how accepting cycle detection algorithms MAP [16] and OWCTY [17] can be accelerated by means of massive parallel processing in order to speedup the LTL model checking procedure.…”

“…Apart from small adjustments of the terminating condition this was the only change necessary for the OWCTY algorithm to be extended to work on multiple CUDA devices. We have duplicated the multi-device experiments from [2] (for MAP algorithm) using our new OWCTY algorithm and we direct the reader to Section 5 to see the comparison of the two implementations.…”