Transport Agent
Dragon takes an innovative approach to HPC networking. Rather than providing a networking library such as MPI or SHMEM, it uses a communication offloading approach where the user application sends communication requests to a separate executable, called the transport agent. The transport agent receives communication requests and handles the actual network communication, all facilitated through Dragon Channels, the core communication layer for Dragon (see Fig. 19).
The use of a transport agent rather than a library provides a number of benefits. Maybe foremost is the ability for processes to come and go efficiently, without any need to initialize a communication library for each new process, since transport agents persist throughout the duration of a job. Another advantage is the ability to aggregate small message across an entire node, rather than only across a single process. This allows for much greater aggregation of small messages, especially as node sizes grow.
TCP Transport
High Speed Transport Agent (HSTA)
Dragon’s High Speed Transport Agent, or HSTA, is an RDMA-based solution for off-node communication of distributed Python multiprocessing applications. HSTA provides a high bandwidth, high message rate and low latency communication framework that aims to make it easy to obtain high network performance via a high-level programming interface. Other popular HPC communication frameworks, such as MPI or SHMEM, generally require low level programming and expert knowledge of both the communication API and the network hardware to obtain optimal performance. Network communication is implicit in applications and workflows using the Dragon ecosystem, allowing HSTA to optimize network communication behind the scenes.
Fig. 19 Dragon Channels off-node communication via the HSTA transport agent