An open-source gymnasium for machine studying assisted laptop structure design – Google Research Blog

Computer Architecture analysis has a protracted historical past of creating simulators and instruments to guage and form the design of laptop programs. For instance, the SimpleScalar simulator was launched within the late Nineteen Nineties and allowed researchers to discover varied microarchitectural concepts. Computer structure simulators and instruments, corresponding to gem5, DRAMSys, and lots of extra have performed a big position in advancing laptop structure analysis. Since then, these shared assets and infrastructure have benefited trade and academia and have enabled researchers to systematically construct on one another’s work, resulting in important advances within the subject.

Nonetheless, laptop structure analysis is evolving, with trade and academia turning in direction of machine studying (ML) optimization to satisfy stringent domain-specific necessities, corresponding to ML for laptop structure, ML for TinyML acceleration, DNN accelerator datapath optimization, reminiscence controllers, energy consumption, safety, and privateness. Although prior work has demonstrated the advantages of ML in design optimization, the shortage of robust, reproducible baselines hinders honest and goal comparability throughout totally different strategies and poses a number of challenges to their deployment. To guarantee regular progress, it’s crucial to grasp and deal with these challenges collectively.

To alleviate these challenges, in “ArchGym: An Open-Source Gymnasium for Machine Learning Assisted Architecture Design”, accepted at ISCA 2023, we launched ArchGym, which incorporates a wide range of laptop structure simulators and ML algorithms. Enabled by ArchGym, our outcomes point out that with a sufficiently massive variety of samples, any of a various assortment of ML algorithms are able to find the optimum set of structure design parameters for every goal downside; nobody resolution is essentially higher than one other. These outcomes additional point out that deciding on the optimum hyperparameters for a given ML algorithm is important for locating the optimum structure design, however selecting them is non-trivial. We launch the code and dataset throughout a number of laptop structure simulations and ML algorithms.

Challenges in ML-assisted structure analysis

ML-assisted structure analysis poses a number of challenges, together with:

1. For a particular ML-assisted laptop structure downside (e.g., discovering an optimum resolution for a DRAM controller) there isn’t any systematic solution to establish optimum ML algorithms or hyperparameters (e.g., studying fee, warm-up steps, and so on.). There is a wider vary of ML and heuristic strategies, from random stroll to reinforcement studying (RL), that may be employed for design house exploration (DSE). While these strategies have proven noticeable efficiency enchancment over their selection of baselines, it isn’t evident whether or not the enhancements are due to the selection of optimization algorithms or hyperparameters.
   Thus, to make sure reproducibility and facilitate widespread adoption of ML-aided structure DSE, it’s needed to stipulate a scientific benchmarking methodology.
2. While laptop structure simulators have been the spine of architectural improvements, there may be an rising want to handle the trade-offs between accuracy, velocity, and price in structure exploration. The accuracy and velocity of efficiency estimation extensively varies from one simulator to a different, relying on the underlying modeling particulars (e.g., cycle–correct vs. ML–primarily based proxy fashions). While analytical or ML-based proxy fashions are nimble by advantage of discarding low-level particulars, they often endure from excessive prediction error. Also, as a consequence of business licensing, there could be strict limits on the variety of runs collected from a simulator. Overall, these constraints exhibit distinct efficiency vs. pattern effectivity trade-offs, affecting the selection of optimization algorithm for structure exploration.
   It is difficult to delineate the way to systematically evaluate the effectiveness of assorted ML algorithms below these constraints.
3. Finally, the panorama of ML algorithms is quickly evolving and a few ML algorithms want information to be helpful. Additionally, rendering the end result of DSE into significant artifacts corresponding to datasets is crucial for drawing insights in regards to the design house.
   In this quickly evolving ecosystem, it’s consequential to make sure the way to amortize the overhead of search algorithms for structure exploration. It is just not obvious, nor systematically studied the way to leverage exploration information whereas being agnostic to the underlying search algorithm.

ArchGym design

ArchGym addresses these challenges by offering a unified framework for evaluating totally different ML-based search algorithms pretty. It includes two foremost parts: 1) the ArchGym setting and a couple of) the ArchGym agent. The setting is an encapsulation of the structure price mannequin — which incorporates latency, throughput, space, power, and so on., to find out the computational price of operating the workload, given a set of architectural parameters — paired with the goal workload(s). The agent is an encapsulation of the ML algorithm used for the search and consists of hyperparameters and a guiding coverage. The hyperparameters are intrinsic to the algorithm for which the mannequin is to be optimized and may considerably affect efficiency. The coverage, however, determines how the agent selects a parameter iteratively to optimize the goal goal.

Notably, ArchGym additionally features a standardized interface that connects these two parts, whereas additionally saving the exploration information because the ArchGym Dataset. At its core, the interface entails three foremost indicators: {hardware} state, {hardware} parameters, and metrics. These indicators are the naked minimal to determine a significant communication channel between the setting and the agent. Using these indicators, the agent observes the state of the {hardware} and suggests a set of {hardware} parameters to iteratively optimize a (user-defined) reward. The reward is a operate of {hardware} efficiency metrics, corresponding to efficiency, power consumption, and so on. 

ArchGym includes two foremost parts: the ArchGym setting and the ArchGym agent. The ArchGym setting encapsulates the price mannequin and the agent is an abstraction of a coverage and hyperparameters. With a standardized interface that connects these two parts, ArchGym offers a unified framework for evaluating totally different ML-based search algorithms pretty whereas additionally saving the exploration information because the ArchGym Dataset.

ML algorithms may very well be equally favorable to satisfy user-defined goal specs

Using ArchGym, we empirically show that throughout totally different optimization targets and DSE issues, not less than one set of hyperparameters exists that ends in the identical {hardware} efficiency as different ML algorithms. A poorly chosen (random choice) hyperparameter for the ML algorithm or its baseline can result in a deceptive conclusion {that a} explicit household of ML algorithms is best than one other. We present that with ample hyperparameter tuning, totally different search algorithms, even random stroll (RW), are in a position to establish the absolute best reward. However, observe that discovering the suitable set of hyperparameters might require exhaustive search and even luck to make it aggressive.

With a ample variety of samples, there exists not less than one set of hyperparameters that ends in the identical efficiency throughout a spread of search algorithms. Here the dashed line represents the utmost normalized reward. Cloud-1, cloud-2, stream, and random point out 4 totally different reminiscence traces for DRAMSys (DRAM subsystem design house exploration framework).

Dataset building and high-fidelity proxy mannequin coaching

Creating a unified interface utilizing ArchGym additionally permits the creation of datasets that can be utilized to design higher data-driven ML-based proxy structure price fashions to enhance the velocity of structure simulation. To consider the advantages of datasets in constructing an ML mannequin to approximate structure price, we leverage ArchGym’s capacity to log the information from every run from DRAMSys to create 4 dataset variants, every with a distinct variety of information factors. For every variant, we create two classes: (a) Diverse Dataset, which represents the information collected from totally different brokers (ACO, GA, RW, and BO), and (b) ACO solely, which reveals the information collected completely from the ACO agent, each of that are launched together with ArchGym. We practice a proxy mannequin on every dataset utilizing random forest regression with the target to foretell the latency of designs for a DRAM simulator. Our outcomes present that:

1. As we improve the dataset dimension, the common normalized root imply squared error (RMSE) barely decreases.
2. However, as we introduce variety within the dataset (e.g., accumulating information from totally different brokers), we observe 9× to 42× decrease RMSE throughout totally different dataset sizes.

Diverse dataset assortment throughout totally different brokers utilizing ArchGym interface.

The influence of a various dataset and dataset dimension on the normalized RMSE.

The want for a community-driven ecosystem for ML-assisted structure analysis

While, ArchGym is an preliminary effort in direction of creating an open-source ecosystem that (1) connects a broad vary of search algorithms to laptop structure simulators in an unified and easy-to-extend method, (2) facilitates analysis in ML-assisted laptop structure, and (3) kinds the scaffold to develop reproducible baselines, there are a number of open challenges that want community-wide help. Below we define a few of the open challenges in ML-assisted structure design. Addressing these challenges requires a properly coordinated effort and a group pushed ecosystem.

Key challenges in ML-assisted structure design.

We name this ecosystem Architecture 2.0. We define the important thing challenges and a imaginative and prescient for constructing an inclusive ecosystem of interdisciplinary researchers to deal with the long-standing open issues in making use of ML for laptop structure analysis. If you have an interest in serving to form this ecosystem, please fill out the curiosity survey.

Conclusion

ArchGym is an open supply gymnasium for ML structure DSE and permits an standardized interface that may be readily prolonged to swimsuit totally different use instances. Additionally, ArchGym permits honest and reproducible comparability between totally different ML algorithms and helps to determine stronger baselines for laptop structure analysis issues.

We invite the pc structure group in addition to the ML group to actively take part within the improvement of ArchGym. We imagine that the creation of a gymnasium-type setting for laptop structure analysis can be a big step ahead within the subject and supply a platform for researchers to make use of ML to speed up analysis and result in new and revolutionary designs.

Acknowledgements

This blogpost relies on joint work with a number of co-authors at Google and Harvard University. We wish to acknowledge and spotlight Srivatsan Krishnan (Harvard) who contributed a number of concepts to this challenge in collaboration with Shvetank Prakash (Harvard), Jason Jabbour (Harvard), Ikechukwu Uchendu (Harvard), Susobhan Ghosh (Harvard), Behzad Boroujerdian (Harvard), Daniel Richins (Harvard), Devashree Tripathy (Harvard), and Thierry Thambe (Harvard).  In addition, we might additionally wish to thank James Laudon, Douglas Eck, Cliff Young, and Aleksandra Faust for his or her help, suggestions, and motivation for this work. We would additionally wish to thank John Guilyard for the animated determine used on this put up. Amir Yazdanbakhsh is now a Research Scientist at Google DeepMind and Vijay Janapa Reddi is an Associate Professor at Harvard.