Mapping applications onto FPGA-centric clusters
MetadataShow full item record
High Performance Computing (HPC) is becoming increasingly important throughout science and engineering as ever more complex problems must be solved through computational simulations. In these large computational applications, the latency of communication between processing nodes is often the key factor that limits performance. An emerging alternative computer architecture that addresses the latency problem is the FPGA-centric cluster (FCC); in these systems, the devices (FPGAs) are directly interconnected and thus many layers of hardware and software are avoided. The result can be scalability not currently achievable with other technologies. In FCCs, FPGAs serve multiple functions: accelerator, network interface card (NIC), and router. Moreover, because FPGAs are configurable, there is substantial opportunity to tailor the router hardware to the application; previous work has demonstrated that such application-aware configuration can effect a substantial improvement in hardware efficiency. One constraint of FCCs is that it is convenient for their interconnect to be static, direct, and have a two or three dimensional mesh topology. Thus, applications that are naturally of a different dimensionality (have a different logical topology) from that of the FCC must be remapped to obtain optimal performance. In this thesis we study various aspects of the mapping problem for FCCs. There are two major research thrusts. The first is finding the optimal mapping of logical to physical topology. This problem has received substantial attention by both the theory community, where topology mapping is referred to as graph embedding, and by the High Performance Computing (HPC) community, where it is a question of process placement. We explore the implications of the different mapping strategies on communication behavior in FCCs, especially on resulting load imbalance. The second major research thrust is built around the hypothesis that applications that need to be remapped (due to differing logical and physical topologies) will have different optimal router configurations from those applications that do not. For example, due to remapping, some virtual or physical communication links may have little occupancy; therefore fewer resources should be allocated to them. Critical here is the creation of a new set of parameterized hardware features that can be configured to best handle load imbalances caused by remapping. These two thrusts form a codesign loop: certain mapping algorithms may be differentially optimal due to application-aware router reconfiguration that accounts for this mapping. This thesis has four parts. The first part introduces the background and previous work related to communication in general and, in particular, how it is implemented in FCCs. We build on previous work on application-aware router configuration. The second part introduces topology mapping mechanisms including those derived from graph embeddings and a greedy algorithm commonly used in HPC. In the third part, topology mappings are evaluated for performance and imbalance; we note that different mapping strategies lead to different imbalances both in the overall network and in each node. The final part introduces reconfigure router design that allocates resources based on different imbalance situations caused by different mapping behaviors.