| Description |
High-performance supercomputers on the Top500 list are commonly designed around commodity CPUs. Most of the codes executed on these machines are message-passing codes using the message-passing toolkit (MPI). Thus it makes sense to look at these machines from a holistic systems architecture perspective and consider optimizations to commodity processors that make them more efficient in message-passing architectures. Described herein is a new User-Level Notification (ULN) architecture that significantly improves message-passing performance. The architecture integrates a simultaneous multithreaded (SMT) processor with a user-level network interface (NI) that can directly control the execution scheduling of threads on the processor. By allowing the network interface to control the execution of message handling code at the user level, the operating system (OS) related overhead for handling interrupts and user code dispatch related to notifications is eliminated. By using an SMT processor, message handling can be performed in one thread concurrent to user computation in other threads, thus most of the overhead of executing message handlers can be hidden. This dissertation presents measurements showing the OS overheads related to message-passing are significant in modern architectures and describes a new architecture that significantly reduces these overheads. On a communication-intensive real-world application, the ULN architecture provides a 50.9% performance improvement over a more traditional OS-based NIC and a 5.29-31.9% improvement over a best-of-class user-level NIC due to the user-level notifications. |
