Adapting a message-driven parallel application to GPU-accelerated clusters