CoPTA: Contiguous Pattern Speculating TLB Architecture

Abstract

With the growing size of real-world datasets running on CPUs, address translation has become a significant performance bottleneck. To translate virtual addresses into physical addresses, modern operating systems perform several levels of page table walks (PTWs) in memory. Translation look-aside buffers (TLBs) are used as caches to keep recently used translation information. However, as datasets increase in size, both the TLB miss rate and the overhead of PTWs worsen, causing severe performance bottlenecks. Using a diverse set of workloads, we show the PTW overhead consumes an average of 20% application execution time.

In this paper, we propose CoPTA, a technique to speculate the memory address translation upon a TLB miss to hide the PTW latency. Specifically, we show that the operating system has a tendency to map contiguous virtual memory pages to contiguous physical pages. Using a real machine, we show that the Linux kernel can automatically defragment physical memory and create larger chunks for contiguous mapping, particularly when transparent huge page support is enabled. Based on this observation, we devise a speculation mechanism that finds nearby entries present in the TLB upon a miss and predicts the address translation of the missed address assuming contiguous address allocation. This allows CoPTAto speculatively execute instructions without waiting for the PTW to complete. We run the PTW in parallel, compare the speculated and the translated physical addresses, and flush the pipeline upon a wrong speculation with similar techniques used for handling branch mispredictions.

We comprehensively evaluate our proposal using benchmarks from three suites: SPEC CPU 2006 for server-grade applications, GraphBIG for graph applications, and the NAS benchmark suite for scientific applications. Using a trace-based simulation, we show an average address prediction accuracy of 82% across these workloads resulting in a 16% performance improvement.