Security of High-Performance Messaging Layers on Programmable Network Interface Architectures

摘要

This paper discusses security concerns introduced by high-performance messaging layers designed for distributed High Performance Computing (HPC) environments. In the recent past, computationally intensive scientific applications have begun to find high-speed Networks of Workstations (NOWs) as an economically attractive alternative to parallel computers. Conventional distributed systems have been studied and modeled for security. However, these models are not directly applicable to HPC systems primarily due to the design goals of the underlying system software, such as high-bandwidth and low-overhead communication and scalability. New security concerns also arise from the programmability of the network interfaces of the underlying high-speed network technologies. This paper presents guidelines to improve the assurance of typical high-performance messaging protocol layers operating in a System Area Network (SAN). This paper identifies that an integral part of the security of any performance critical system is providing immunity from performance-degrading attacks. A secure messaging system should minimize threats to the availability of limited and shared resources in order to preserve the system's performance. The performance goals of HPC systems severely restrict the choice of usable security mechanisms, because of the inevitable processing overhead introduced. In view of this restriction, a trade-off model has been suggested to achieve the desired balance between performance and assurance.