Freya PAGE: First Optimal Time Complexity for Large-Scale Nonconvex Finite-Sum Optimization with Heterogeneous Asynchronous Computations

In practical distributed systems, workers are typically not homogeneous, and due to differences in hardware configurations and network conditions, can have highly varying processing times. We consider smooth nonconvex finite-sum (empirical risk minimization) problems in this setup and introduce a new parallel method, Freya PAGE, designed to handle arbitrarily heterogeneous and asynchronous computations. By being robust to "stragglers" and adaptively ignoring slow computations, Freya PAGE offers significantly improved time complexity guarantees compared to all previous methods, including Asynchronous SGD, Rennala SGD, SPIDER, and PAGE, while requiring weaker assumptions. The algorithm relies on novel generic stochastic gradient collection strategies with theoretical guarantees that can be of interest on their own, and may be used in the design of future optimization methods. Furthermore, we establish a lower bound for smooth nonconvex finite-sum problems in the asynchronous setup, providing a fundamental time complexity limit. This lower bound is tight and demonstrates the optimality of Freya PAGE in the large-scale regime, i.e., when $\sqrt{m} \geq n$, where $n$ is # of workers, and $m$ is # of data samples.

Paper

References (33)

Scroll for more · 21 remaining

Similar papers

Peer review

Reviewer wr2A6/10 · confidence 3/52024-06-27

Summary

In this study, the authors developed a new distributed optimization algorithm, Freya Page method, for the non-convex finite-sum optimization problem. The authors provided the corresponding iteration and time complexity bounds. In addition, a lower bound on the time complexity is also provided.

Strengths

The paper is well-written and easy to follow. The developed Algorithms 2-3 are novel and lead to useful time complexity bounds. The results should be interesting to audiences in distributed optimization and machine learning fields. Due to the time limit, I did not check the proofs in the appendix. But I feel that the proofs should be correct, except that of Theorem 10, which I have a concern about.

Weaknesses

I did not see major weaknesses. But I am a little confused by the lower bound in Theorem 10 and the claims after the theorem. I have also included a few comments for the authors' consideration.

Questions

- I wonder if the results can be extended to the case when the computation time for each gradient $\nabla f_j$ is different for different $j$. In other words, can we replace $\tau_{i}$ in Line 31 with $\tau_{i, j}$, which may be different for different $j$? - Line 118: It would be better if the authors can provide more details on the choice of inverse proportion to $\tau_i$. I feel that this choice can guarantee that the worst-case running time is the same for all workers? - In Section 1, the authors claimed that Theorem 4 allows $\tau_i = \infty$ for all $i$. But I feel that the authors are referring to Theorem 5? Also, I think Theorem 5 (as well as Theorems 1-2) requires $\tau_i$ to be finite for some $i$ so that $t^*$ is finite? It seems that this is supported by Example 3, where the current analysis is not able to provide an upper bound on the time complexity. - For the time complexity in Theorem 4, I think the authors are computing the *expected* time complexity. This information is not mentioned in the paper and I feel it might be better to add a formal definition of the expected time complexity in the paper. - It seems that the authors defined the notion of $t^*(S)$ in Definition 3 but did not used it in the following theorems. - Line 200: I am not sure why the monotony of the two terms enable the application of the bi-section method. - Line 294: I am a little confused when the authors mentioned that (10) is less than or equal to (12). It seems that (12) should be a lower bound on the complexity (up to a constant) and the achieved time complexity (10) should not be smaller than it?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

See my comments in the Questions section.

Reviewer 6ZMd5/10 · confidence 3/52024-07-11

Summary

The paper introduces Freya PAGE, a novel parallel optimization method designed for distributed systems where computational resources vary in capability and speed. This method addresses the challenge of stragglers in asynchronous environments by adopting a strategy that adaptively ignores slower computations, thereby significantly improving the time complexity over previous methods like Asynchronous SGD and PAGE. The paper claims that Freya PAGE achieves optimal time complexity under weaker assumptions and also presents a theoretical analysis that establishes a lower bound for the time complexity in such settings. The approach is particularly effective in large-scale scenarios where the number of data samples is greater than the square root of the number of workers, thus demonstrating the algorithm's suitability for practical large-scale machine learning tasks.

Strengths

1. The problem tackled is relevant to current large-scale machine learning applications, making the findings applicable and valuable to both academics and practitioners working with distributed computing environments. 2. Freya PAGE introduces a novel approach to handling heterogeneous and asynchronous computational environments in distributed systems. Its ability to adaptively ignore slower computations is a significant advancement over traditional methods. 3. The paper demonstrates that Freya PAGE achieves an optimal time complexity, which is a significant improvement compared to existing methods like asynchronous SGD.

Weaknesses

1. Determining the optimal parameters $S$ and $p$ as outlined in Theorem 7 requires knowledge of unknown $\tau_i$ and solving an optimization problem, which can be challenging or even impractical. 2. The paper lacks sufficient experimental results on more complex machine learning tasks, including comparision with cutting-edge distributed learning algorithms. 3. The algorithm does not address the non-iid data senarios, workers may access data from varying distributions.

Questions

How do the workers operate asynchronously? And why do the authors assert that Freya PAGE combines synchrony and asynchrony?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors have partially addressed the limitations.

Authorsrebuttal2024-08-12

Official Comment by Authors

Thank you for the question. Indeed, in Line 304, we say that Freya SGD is closely related to Rennala SGD [1]. We consider Freya SGD to get the time complexity of the finite-sum optimization with a *non-variance reduced method.* Why is it important? First, such a method can be crucial and has a better convergence rate in the interpolation regime [2,3]. Second, while the dependence of Freya SGD on $\varepsilon$ is worse compared to Freya PAGE, the former method does not *explicitly* depend on $m$ (compare (10) vs the last formula in Theorem 25), which can be beneficial in some regimes. In total, Freya SGD provides a new time complexity that was not considered in the literature, which we believe is important for future work and for a general understanding of asynchronous methods. Let us explain the main differences between Freya SGD and Rennala SGD. The time complexity of Freya SGD is equal to $$\frac{\delta^0 L_-}{\varepsilon} \min_{j \in [n]} \left(\left(\sum_{i=1}^j \frac{1}{\tau_i}\right)^{-1} \left(\frac{L_{\max}}{\varepsilon} \left(\delta^0 + \Delta^*\right) + j \right)\right).$$ Unlike Rennala SGD, this complexity does not depend on the noise $\sigma^2$ of stochastic gradients, which is a major advantage of Freya SGD. Another important difference between Rennala SGD and Freya SGD is the choice of the number of stochastic gradients that methods calculate in every iteration. In [1], the authors choose $S \approx \frac{\sigma^2}{\varepsilon},$ while we prove that the optimal choice of $S$ that minimizes the upper bound in Theorem 25 is $\approx \frac{L_{\max}}{\varepsilon} \left(\delta^0 + \Delta^*\right).$ Moreover, the convergence rate of Rennala SGD relies on the theory from [4], while Freya SGD uses the analysis with a different proof technique from [5]. Beyond the main contributions related to Freya PAGE and the lower bound, we think that the analysis of Freya SGD is the cherry on top, offering deeper insights into time complexities in the finite-sum setting with heterogeneous asynchronous computations. We hope we have answered the question. If you need more information, we'd be happy to provide further details. [1] Tyurin and Richtárik, “Optimal Time Complexities of Parallel Stochastic Optimization Methods under a Fixed Computation Model”, in NeurIPS 2023. [2] M. Schmidt and N. L. Roux. Fast convergence of stochastic gradient descent under a strong growth condition. arXiv preprint arXiv:1308.6370, 2013. [3] S. Ma, R. Bassily, and M. Belkin. The power of interpolation: Understanding the effectiveness of SGD in modern overparametrized learning. In International Conference on Machine Learning, pages 3325–3334. PMLR, 2018. [4] Ghadimi, S. and Lan, G. (2013). Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23(4):2341–2368. [5] A. Khaled and P. Richtárik. Better theory for SGD in the nonconvex world. Transactions on Machine Learning Research, 2022.

Authorsrebuttal2024-08-13

Optimality

We wish to re-iterate that despite decades of research on SGD (SGD was studied at least since 1951; Robbins and Monroe) and on asynchronous methods (1986; Tsitsiklis et al), FreyaPAGE is the first provably optimal asynchronous SGD method for minimizing finite sums in the smooth nonconvex data-homogeneous regime. We are biased of course, but we believe such a result should command a substantially higher score than 5, despite the three shortcomings mentioned by the reviewer. We addressed the first weakness in our rebuttal - please check our response. We view the second weakness as minor since our work is of a theoretical nature. We believe theoretical works should be judged on the basis of the strength of the theory, just like empirical works should be judged based on the experimental results. Nevertheless, we will include several additional well-designed experiments in the camera-ready version of the paper. Lastly, we acknowledge the third weakness. Having said that, our work is a first step to considering more challenging scenarios in the future, by us or others in the community. Of course, we understand the reviewer may have a different view. Nevertheless, we thought it might be useful to say how we feel about the theoretical importance of our result. Thanks again for your review and for considering our rebuttal and responses! Authors

Reviewer QBDK8/10 · confidence 3/52024-07-12

Summary

The paper introduces Freya PAGE, a new parallel method for large-scale nonconvex finite-sum optimization in heterogeneous and asynchronous computational environments. Freya PAGE, specifically addresses the variability in processing times across different workers due to hardware and network differences. By being robust to "stragglers" and adaptively ignoring slow computations, Freya PAGE improves time complexity guarantees compared to existing methods. Additionally, the paper proves a lower bound for smooth nonconvex finite-sum problems in asynchronous setups.

Strengths

Freya PAGE is a novel parallel optimization method specifically designed to handle heterogeneous and asynchronous computations. This appears to be a significant advancement and addresses real-world challenges in distributed systems. By addressing the variability in processing times across different workers due to hardware and network differences, the method is highly relevant to practical distributed systems used in large-scale machine learning tasks. This method shows improved time complexity guarantees compared to existing methods. The paper also provides a strong theoretical framework in general. Specifically, proving lower bound for smooth nonconvex finite-sum problems in asynchronous setups shows the optimality of Freya PAGE in large-scale regimes. The stochastic gradient collection strategies ComputeGradient and ComputeBatchDifference introduced in this paper is innovative. Overall, this is a very well written paper and easy to follow and understand.

Weaknesses

The empirical results of this paper are primarily focused on synthetic quadratic optimization tasks and logistic regression problems. More diverse and extensive experimentation on a wider range of real-world datasets and applications would strengthen the practical validation. The algorithm's performance might vary significantly for different configurations of worker speeds and hardware setups. The variance in performance could be an issue in highly dynamic environments.

Questions

If you have tested Freya PAGE in real-world scenarios, can you share the results and insights? How does Freya PAGE specifically determine which computations to ignore or adaptively manage in the presence of stragglers? Can this mechanism be fine-tuned for different distributed system configurations?

Rating

8

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

The performance of Freya PAGE might vary significantly with different configurations of worker speeds and hardware setups. This variance could be problematic in environments with highly variable computational resources. The impact of communication overhead in distributed systems with significant network latency, is not extensively discussed.

Reviewer MLsi6/10 · confidence 3/52024-07-13

Summary

This paper considers the problem of minimizing a finite sum of non-convex objective terms with multiple computational units that calculate gradient oracles. It introduces two efficient computational subroutines for the main steps of the PAGE algorithm, and provides novel bounds on the real time of computation assuming a heterogeneous computational scenario. For this, the paper gives bounds on the execution time of these subroutines and combines them with the previously known iteration complexity of PAGE. Next, the paper considers multiple variations: it optimizes the hyperparameters of the PAGE algorithm in different regimes, and also extends the results to a dynamic scenario. Finally, it provides a lower bound on the overall execution time and in this way establishes that their strategy is optimal in large-scale regimes.

Strengths

The paper is well written and has an excellent presentation of the subject. It provides an extensive theoretical study of a novel gradient management strategy in the PAGE algorithms, and achieves tight bounds on the convergence rate.

Weaknesses

I do not have any major concern about this work. The only limitation for me is that from a practical standpoint, this algorithm might not be optimal as it may introduce a significantly higher communication overhead. This is because each gradient is now required to be sent in an individual message.

Questions

Compared to the results in the appendix, the stated theorems are generally simplified for better comprehension, but some confusion is also introduced. In particular, I am not sure which results are deterministic and which results are in terms of statistical expectation or high-probability. Could you clarify? In the statement of Theorem 4, you assume two bounds on m, where one of them seem to be stronger than the other. Why do you keep both?

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

No limitations or potential negative societal impact.

Reviewer QBDK2024-08-10

On the rebuttal

I thank the authors for their responses. I have read the responses carefully and I have no further questions. I am inclined to retain my original scores.

Reviewer 6ZMd2024-08-12

Thank the authors for addressing my initial concerns. I notice that Freya SGD is quite similar to Rennala SGD [1] mentioned in the manuscript. While Freya SGD is designed for finite-sum optimization problems and Rennala SGD addresses stochastic programming, their adaptation from one context to another appears to be straightforward. Could the authors detail the distinct algorithmic and technical challenges that Freya SGD specifically addresses? This comparison would help delineate the unique contributions inherent in Freya SGD, particularly in how it navigates the complexities of finite-sum optimization compared to the broader scope of stochastic programming tackled by Rennala SGD. [1] Tyurin and Richtárik, “Optimal Time Complexities of Parallel Stochastic Optimization Methods under a Fixed Computation Model”, in NeurIPS 2023.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC