Mold Linker Paper Claims Up to 112x Speedup Over GNU ld, Accepted to ASPLOS 2027
A new paper on the mold linker reports it is 2.4-16.1x faster than lld and up to 112x faster than GNU ld, and has been accepted to ASPLOS 2027.
Overview
A paper describing mold, an open-source Unix/Linux linker, reports that it is “2.4-16.1x faster than the state-of-the-art lld linker, and up to 112x faster than the traditional GNU ld,” according to the paper’s abstract on arXiv. The paper, titled “mold: A Massively Parallel Linker,” was submitted by author Rui Ueyama and has been accepted to ASPLOS 2027, the Association for Computing Machinery’s conference on Architectural Support for Programming Languages and Operating Systems.
What We Know
Linking — the final stage of a software build that combines compiled object files into an executable or shared library — has lagged behind advances in parallel compilation, according to the paper. “Despite extensive research on compilation, the linking step has remained largely neglected. Yet for large programs, linking alone can take tens of seconds, becoming the bottleneck in the edit-compile-debug cycle that developers repeat dozens of times a day,” the paper states, as published on arXiv.
The paper illustrates the problem with a real-world example: “A debug build of TensorFlow, for example, comprises 24 GiB of object files, which the linker combines into a single 9.9 GiB shared library. Linking such a program with the state-of-the-art LLVM lld takes 52 seconds on a 64-core machine, leaving most of its cores idle,” according to the paper. Benchmarks in the paper show mold completing the same TensorFlow link in roughly a tenth of that time, delivering the paper’s reported 16.1x speedup over lld for that workload.
The paper’s abstract describes mold’s approach as “a Unix/Linux linker that applies data parallelism systematically across the entire linking pipeline,” adding that the design “decouples” symbol resolution and archive processing — constraints the authors say prevent existing linkers from scaling across CPU cores. The paper also reports an ablation study, concluding that “no single optimization dominates; the speedup comes from the cumulative effect of parallelizing all passes,” per the abstract.
Benchmarks described in the paper’s evaluation section were run on “an AMD Ryzen Threadripper 7980X (64 cores, 128 threads), 384 GiB of DDR5 RAM, running Linux 6.17 on Ubuntu 24.04,” according to the paper. On Chromium, the paper’s benchmark table shows mold completing a debug-build link in 1.89 seconds versus 13.24 seconds for lld.
Ueyama is also the original author of LLVM’s lld, the linker mold’s benchmarks are measured against. Mold’s GitHub repository describes the project as “a faster drop-in replacement for existing Unix linkers,” adding that it is “several times quicker than the LLVM lld linker, the second-fastest open-source linker, which I initially developed a few years ago.”
What We Don’t Know
The paper does not disclose which companies or open-source projects currently use mold in production, nor does it specify a general-availability timeline tied to the ASPLOS 2027 conference presentation. The submitted version is a preprint; any changes made during ASPLOS’s peer-review process before final publication are not yet reflected in the current text.
Analysis
Linkers are an unglamorous but unavoidable part of every compiled-language build, and their runtime compounds across the thousands of edit-compile-debug cycles a large engineering team runs daily. A linker that turns a nearly minute-long wait — as described in the paper’s TensorFlow example — into a few seconds addresses a cost that is easy to overlook until it is measured in aggregate developer time. That the paper’s author is also the creator of lld, the very tool mold is benchmarked against, underscores how much of the recent work on build-tool performance is happening within a small community of specialists who have already built the previous generation of “fast” tooling and are now trying to outdo it.