Efficient Physical Page Migrations in Shared Virtual Memory Reconfigurable Computing Systems

Abstract

Shared Virtual Memory (SVM) can considerably simplify the application development for FPGA-accelerated computers, as it allows the seamless passing of virtually addressed pointers across the hardware/software boundary. Especially applications operating on complex pointer-based data structures can profit from this approach, as SVM can often avoid having to copy the entire data to FPGA memory, while performing pointer relocations in the process. Many FPGA-accelerated computers, especially in a data center setting, employ PCIe-attached boards that have FPGA-local memory in the form of on-chip HBM or on-board DRAM. Accesses to this local memory are much faster than going to the host memory via PCIe. Thus, even in the presence of SVM, it is desirable to be able to move the physical memory pages holding frequently accessed data closest to the compute unit that is operating on them. This capability is called physical page migration. The main contribution of this work is an open-source framework which provides SVM with physical page migration capabilities to PCIe-attached FPGA cards. We benchmark both fully automatic on-demand and user-managed explicit migration modes, and show that for suitable use-cases, the performance of migrations cannot just match that of conventional DMA copy-based accelerator operations, but may even exceed it by overlapping computations and migrations.