Trident: Harnessing Architectural Resources for All Page Sizes in x86 Processors

Intel and AMD's processors have long supported two large page sizes -- 1GB and 2MB, to reduce address translation overheads for big data applications. However, previous works on large pages have focused on 2MB pages, partly due to a lack of evidence on the usefulness of 1GB pages to real-world applications. Consequently, micro-architectural resources devoted to 1GB pages have gone underutilized for a decade. In this work, we first quantitatively demonstrate where 1GB pages can be valuable, especially when employed in conjunction with 2MB pages. We find that inadequate system software enablement is to blame for the under-utilization of 1GB pages. We, thus, design and implement Trident in Linux to fully harness micro-architectural resources devoted for all page sizes in the current x86 hardware by transparently allocating 1GB, 2MB, and 4KB pages as suitable at runtime. Trident speeds up eight memory-intensive applications by 18%, on average, over Linux's use of 2MB pages. We also propose TridentPV, an extension to \trident{} that effectively virtualizes 1GB pages via copy-less promotion and compaction in the guest OS. Overall, this paper shows that adequate software enablement can bring practical relevance to even GB-sized pages, in turn motivating micro-architects to continue investing/innovating in large pages. This artifact provides source of Trident (modified Linux kernel v4.17.3) and evaluated benchmark binaries with their input files where appropriate. The measurement infrastructure is provided through bash and python scripts which allow reproducing experimental results on an Intel Skylake machine with 18 cores (36 threads) and 192GB of main memory