Source code for our ASPLOS 2023 paper, "MiniMalloc: A Lightweight Memory Allocator for Hardware-Accelerated Machine Learning."
An increasing number of deep learning workloads are being supported by hardware acceleration. In order to unlock the maximum performance of a hardware accelerator, a machine learning model must first be carefully mapped onto its various internal components by way of a compiler. One especially important problem faced by a production-class compiler is that of memory allocation, whereby a set of buffers with predefined lifespans are mapped onto offsets in global memory. Since this allocation is performed statically, the compiler has the freedom to place buffers strategically, but must nevertheless wrestle with a combinatorial explosion in the number of assignment possibilities.
MiniMalloc is a state-of-the-art algorithm designed specifically for static memory allocation that uses several novel search techiques in order to solve such problems efficiently and effectively.
A key insight motivating our methodology is the discovery of a specific category of solutions -- which we call canonical solutions -- that correspond to the members of an algebraic lattice:
By limiting our exploration to the subset of canonical solutions, we can dramatically reduce the size of the search space while simultaneously ensuring that our algorithm remains sound and complete. We also employ a new spatial inference technique that takes advantage of this special structure, allowing our solver to backtrack much earlier than otherwise possible. Finally, we implement a new mechanism for detecting and eliminating dominated solutions from consideration.
$ git clone --recursive git@github.com:google/minimalloc.git && \
cd minimalloc && cmake -DCMAKE_BUILD_TYPE=Release && make
id,lower,upper,size
b1,0,3,4
b2,3,9,4
b3,0,9,4
b4,9,21,4
b5,0,21,4
$ ./minimalloc --capacity=12 --input=benchmarks/examples/input.12.csv --output=output.12.csv
id,lower,upper,size,offset
b1,0,3,4,8
b2,3,9,4,8
b3,0,9,4,4
b4,9,21,4,4
b5,0,21,4,0
@inproceedings{Moffitt2023,
title = {{MiniMalloc}: A Lightweight Memory Allocator for Hardware-Accelerated Machine Learning},
booktitle = {Proceedings of the 28th International Conference on Architectural Support for Programming Languages and Operating Systems},
volume = {4},
pages = {238--252},
author = {Moffitt, Michael D.},
year = {2023},
series = {ASPLOS 2023},
url = {https://doi.org/10.1145/3623278.3624752},
doi = {10.1145/3623278.3624752}
}
This is not an officially supported Google product.