Pc Structure analysis has an extended historical past of creating simulators and instruments to guage and form the design of pc techniques. For instance, the SimpleScalar simulator was launched within the late Nineties and allowed researchers to discover numerous microarchitectural concepts. Pc structure simulators and instruments, corresponding to gem5, DRAMSys, and plenty of extra have performed a major function in advancing pc 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 vital advances within the discipline.

Nonetheless, pc 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 pc structure, ML for TinyML acceleration, DNN accelerator datapath optimization, reminiscence controllers, energy consumption, safety, and privateness. Though prior work has demonstrated the advantages of ML in design optimization, the dearth of sturdy, reproducible baselines hinders truthful and goal comparability throughout totally different strategies and poses a number of challenges to their deployment. To make sure regular progress, it’s crucial to grasp and deal with these challenges collectively.

To alleviate these challenges, in “ArchGym: An Open-Supply Gymnasium for Machine Studying Assisted Structure Design”, accepted at ISCA 2023, we launched ArchGym, which incorporates a wide range of pc structure simulators and ML algorithms. Enabled by ArchGym, our outcomes point out that with a sufficiently giant 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 crucial for locating the optimum structure design, however selecting them is non-trivial. We launch the code and dataset throughout a number of pc 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 selected ML-assisted pc structure downside (e.g., discovering an optimum resolution for a DRAM controller) there is no such thing as a systematic solution to determine optimum ML algorithms or hyperparameters (e.g., studying charge, warm-up steps, and many others.). There’s a wider vary of ML and heuristic strategies, from random stroll to reinforcement studying (RL), that may be employed for design area exploration (DSE). Whereas these strategies have proven noticeable efficiency enchancment over their selection of baselines, it’s not 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 obligatory to stipulate a scientific benchmarking methodology.
2. Whereas pc structure simulators have been the spine of architectural improvements, there may be an rising want to deal with the trade-offs between accuracy, velocity, and value in structure exploration. The accuracy and velocity of efficiency estimation broadly varies from one simulator to a different, relying on the underlying modeling particulars (e.g., cycle–correct vs. ML–primarily based proxy fashions). Whereas analytical or ML-based proxy fashions are nimble by advantage of discarding low-level particulars, they often undergo from excessive prediction error. Additionally, on account of business licensing, there will be strict limits on the variety of runs collected from a simulator. Total, these constraints exhibit distinct efficiency vs. pattern effectivity trade-offs, affecting the selection of optimization algorithm for structure exploration.
   It’s difficult to delineate how one can systematically evaluate the effectiveness of varied ML algorithms below these constraints.
3. Lastly, the panorama of ML algorithms is quickly evolving and a few ML algorithms want knowledge to be helpful. Moreover, rendering the end result of DSE into significant artifacts corresponding to datasets is crucial for drawing insights in regards to the design area.
   On this quickly evolving ecosystem, it’s consequential to make sure how one can amortize the overhead of search algorithms for structure exploration. It’s not obvious, nor systematically studied how one can leverage exploration knowledge 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 contains two predominant elements: 1) the ArchGym surroundings and a pair of) the ArchGym agent. The surroundings is an encapsulation of the structure price mannequin — which incorporates latency, throughput, space, power, and many others., 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 might considerably affect efficiency. The coverage, alternatively, determines how the agent selects a parameter iteratively to optimize the goal goal.

Notably, ArchGym additionally features a standardized interface that connects these two elements, whereas additionally saving the exploration knowledge because the ArchGym Dataset. At its core, the interface entails three predominant alerts: {hardware} state, {hardware} parameters, and metrics. These alerts are the naked minimal to ascertain a significant communication channel between the surroundings and the agent. Utilizing these alerts, 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 many others. 

ArchGym contains two predominant elements: the ArchGym surroundings and the ArchGym agent. The ArchGym surroundings encapsulates the associated fee mannequin and the agent is an abstraction of a coverage and hyperparameters. With a standardized interface that connects these two elements, ArchGym supplies a unified framework for evaluating totally different ML-based search algorithms pretty whereas additionally saving the exploration knowledge because the ArchGym Dataset.

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

Utilizing ArchGym, we empirically exhibit that throughout totally different optimization targets and DSE issues, not less than one set of hyperparameters exists that leads to 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 healthier than one other. We present that with ample hyperparameter tuning, totally different search algorithms, even random stroll (RW), are in a position to determine the absolute best reward. Nevertheless, notice that discovering the precise set of hyperparameters could 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 leads to the identical efficiency throughout a variety of search algorithms. Right 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 area exploration framework).

Dataset development and high-fidelity proxy mannequin coaching

Making 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 guage the advantages of datasets in constructing an ML mannequin to approximate structure price, we leverage ArchGym’s potential to log the information from every run from DRAMSys to create 4 dataset variants, every with a special variety of knowledge factors. For every variant, we create two classes: (a) Various Dataset, which represents the information collected from totally different brokers (ACO, GA, RW, and BO), and (b) ACO solely, which exhibits the information collected completely from the ACO agent, each of that are launched together with ArchGym. We prepare 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. Nevertheless, as we introduce variety within the dataset (e.g., accumulating knowledge from totally different brokers), we observe 9× to 42× decrease RMSE throughout totally different dataset sizes.

Various dataset assortment throughout totally different brokers utilizing ArchGym interface.

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

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

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

Key challenges in ML-assisted structure design.

We name this ecosystem Structure 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 pc structure analysis. In case you are all for 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. Moreover, ArchGym permits truthful and reproducible comparability between totally different ML algorithms and helps to ascertain stronger baselines for pc structure analysis issues.

We invite the pc structure group in addition to the ML group to actively take part within the growth of ArchGym. We consider that the creation of a gymnasium-type surroundings for pc structure analysis could be a major step ahead within the discipline and supply a platform for researchers to make use of ML to speed up analysis and result in new and modern designs.

Acknowledgements

This blogpost relies on joint work with a number of co-authors at Google and Harvard College. We want to acknowledge and spotlight Srivatsan Krishnan (Harvard) who contributed a number of concepts to this venture 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).  As well as, we might additionally wish to thank James Laudon, Douglas Eck, Cliff Younger, and Aleksandra Faust for his or her assist, suggestions, and motivation for this work. We might additionally wish to thank John Guilyard for the animated determine used on this submit. Amir Yazdanbakhsh is now a Analysis Scientist at Google DeepMind and Vijay Janapa Reddi is an Affiliate Professor at Harvard.