Climate science studies the structure and dynamics of Earth's climate system and seeks to understand how climate changes over time, where the data is usually stored in the format of time series, recording the climate features, geolocation, time attributes, etc. Recently, much research attention has been paid to the climate benchmarks. In addition to the most common task of weather forecasting, several pioneering benchmark works are proposed for extending the modality, such as domain-specific applications like tropical cyclone intensity prediction and flash flood damage estimation, or climate statement and confidence level in the format of natural language. To further motivate the artificial intelligence development for climate science, in this paper, we first contribute a multi-modal climate benchmark, i.e., ClimateBench-M, which aligns (1) the time series climate data from ERA5, (2) extreme weather events data from NOAA, and (3) satellite image data from NASA HLS based on a unified spatial-temporal granularity. Second, under each data modality, we also propose a simple but strong generative method that could produce competitive performance in weather forecasting, thunderstorm alerts, and crop segmentation tasks in the proposed ClimateBench-M. The data and code of ClimateBench-M are publicly available at https://github.com/iDEA-iSAIL-Lab-UIUC/ClimateBench-M.

Fine-tuning large language models (LLMs) with low-rank adaptations (LoRAs) has become common practice, often yielding numerous copies of the same LLM differing only in their LoRA updates. This paradigm presents challenges for systems that serve real-time responses to queries that each involve a different LoRA. Prior works optimize the design of such systems but still require continuous loading and offloading of LoRAs, as it is infeasible to store thousands of LoRAs in GPU memory. To mitigate this issue, we investigate the efficacy of compression when serving LoRAs. We propose a method for the joint compression of LoRAs into a shared basis paired with LoRA-specific scaling matrices. We extend our algorithm to learn clusters of LoRAs that are amenable to joint compression, allowing it to scale gracefully to large LoRA collections. Our experiments with up to 1000 LoRAs demonstrate that compressed LoRAs preserve performance while offering major throughput gains in realistic serving scenarios with over a thousand LoRAs, maintaining 80\% of the throughput of serving a single LoRA.

Most popular benchmarks for comparing LLMs rely on a limited set of prompt templates, which may not fully capture the LLMs’ abilities and can affect the reproducibility of results on leaderboards. Many recent works empirically verify prompt sensitivity and advocate for changes in LLM evaluation. In this paper, we consider the problem of estimating the performance distribution across many prompt variants instead of finding a single prompt to evaluate with. We introduce PromptEval, a method for estimating performance across a large set of prompts borrowing strength across prompts and examples to produce accurate estimates under practical evaluation budgets. The resulting distribution can be used to obtain performance quantiles to construct various robust performance metrics (e.g., top 95% quantile or median). We prove that PromptEval consistently estimates the performance distribution and demonstrate its efficacy empirically on three prominent LLM benchmarks: MMLU, BIG-bench Hard, and LMentry; for example, PromptEval can accurately estimate performance quantiles across 100 prompt templates on MMLU with a budget equivalent to two single-prompt evaluations. Moreover, we show how PromptEval can be useful in LLM-as-a-judge and best prompt identification applications.

We present a toolkit for knowledge-based news event analysis and forecasting. The toolkit is powered by a Knowledge Graph (KG) of events curated from structured and unstructured sources of event-related knowledge. The toolkit provides functions for 1) mapping ongoing news headlines to concepts in the KG, 2) retrieval, reasoning, and visualization for causal analysis and forecasting, and 3) extraction of causal knowledge from text documents to augment the KG with additional domain knowledge. Each function has a number of implementations using a wide range of state-of-the-art neuro-symbolic techniques. We show how the toolkit enables building a human-in-the-loop explainable solution for event analysis and forecasting.

We present Akamai DNS, one of the largest authoritative DNS infrastructures in the world, that supports the Akamai content delivery network (CDN) as well as authoritative DNS hosting and DNS-based load balancing services for many enterprises. As the starting point for a significant fraction of the world's Internet interactions, Akamai DNS serves millions of queries each second and must be resilient to avoid disrupting myriad online services, scalable to meet the ever increasing volume of DNS queries, performant to prevent user-perceivable performance degradation, and reconfigurable to react quickly to shifts in network conditions and attacks. We outline the design principles and architecture used to achieve Akamai DNS's goals, relating the design choices to the system workload and quantifying the effectiveness of those designs. Further, we convey insights from operating the production system that are of value to the broader research community.

We consider the problem of proactive retention aware caching in a heterogeneous wireless edge network consisting of mobile users accessing content from a server and associated to one or more edge caches. Our goal is to design a caching policy that minimizes the sum of content storage costs and server access costs over two design variables: the retention time of each cached content and the probability that a user routes content requests to each of its associated caches. We develop a model that captures multiple aspects such as cache storage costs and several capabilities of modern wireless technologies, such as server multicast/unicast transmissions, device multi-path routing, and cache access constraints. We formulate the problem of Proactive Retention Routing Optimization as a non-convex, non-linear mixed-integer program. We prove that it is NP-hard under both multicast/unicast modes-even when the caches have a large capacity and storage costs are linear-and develop greedy algorithms that have provable performance bounds for the case of uncapacitated caches. Finally, we propose heuristics with low computational complexity for the capacitated cache case as well as for the case of convex storage costs. Systematic evaluations based on real-world data demonstrate the effectiveness of our approach, compared to the existing caching schemes.

The training data and models are becoming increasingly large in many deep-learning applications. Large-scale distributed processing is employed to accelerate training. Increasing the number of learners in synchronous and asynchronous stochastic gradient descent presents challenges to convergence and communication performance. We present our hierarchical, bulk-synchronous stochastic gradient algorithm that effectively balances execution time and accuracy for training in deep-learning applications on GPU clusters. It achieves much better convergence and execution time at scale in comparison to asynchronous stochastic gradient descent implementations. When deployed on a cluster of 128 GPUs, our implementation achieves up to 56 times speedups over the sequential stochastic gradient descent with similar test accuracy for our target application.

Parallel and distributed processing is employed to accelerate training for many deep-learning applications with large models and inputs. As it reduces synchronization and communication overhead by tolerating stale gradient updates, asynchronous stochastic gradient descent (ASGD), derived from stochastic gradient descent (SGD), is widely used. Recent theoretical analyses show ASGD converges with linear asymptotic speedup over SGD. Oftentimes glossed over in theoretical analysis are communication overhead and practical learning rates that are critical to the performance of ASGD. After analyzing the communication performance and convergence behavior of ASGD using the Downpour algorithm as an example, we demonstrate the challenges for ASGD to achieve good practical speedup over SGD. We propose a distributed, bulk-synchronous stochastic gradient descent algorithm that allows for sparse gradient aggregation from individual learners. The communication cost is amortized explicitly by a gradient aggregation interval, and global reductions are used instead of a parameter server for gradient aggregation. We prove its convergence and show that it has superior communication performance and convergence behavior over popular ASGD implementations such as Downpour and EAMSGD for deep-learning applications.

We consider the problem of proactive retention aware caching in a heterogeneous wireless edge network consisting of mobile users connected to a server and associated to one or more edge caches. Our goal is to design a caching policy that minimizes the sum of content storage costs and server access transmissions costs over two design variables: the retention time of each cached content and the probability that a user routes content requests to its associated caches. We develop a model that captures multiple aspects such as cache storage costs and several capabilities of modern wireless technologies, such as server multicast/unicast transmissions, device multipath routing, and cache access constraints. We formulate the problem of Proactive Retention Routing Optimization (PRRO) as a non-convex, non-linear mixed-integer program. We prove that it is NP-Hard under both multicast/unicast modes, even when the caches have a large capacity, and develop a greedy algorithm that has provable performance bounds. Finally, we propose a heuristic for the capacitated cache case that has low computational complexity. Systematic evaluations including real data sets demonstrate the effectiveness of our approach, compared to the existing caching schemes.

We examine the quality of social choice mechanisms using a utilitarian view, in which all of the agents have costs for each of the possible alternatives. While these underlying costs determine what the optimal alternative is, they may be unknown to the social choice mechanism; instead the mechanism must decide on a good alternative based only on the ordinal preferences of the agents which are induced by the underlying costs. Due to its limited information, such a social choice mechanism cannot simply select the alternative that minimizes the total social cost (or minimizes some other objective function). Thus, we seek to bound the distortion: the worst-case ratio between the social cost of the alternative selected and the optimal alternative. Distortion measures how good a mechanism is at approximating the alternative with minimum social cost, while using only ordinal preference information. The underlying costs can be arbitrary, implicit, and unknown; our only assumption is that the agent costs form a metric space, which is a natural assumption in many settings. We quantify the distortion of many well-known social choice mechanisms. We show that for both total social cost and median agent cost, many positional scoring rules have large distortion, while on the other hand Copeland and similar mechanisms perform optimally or near-optimally, always obtaining a distortion of at most 5. We also give lower bounds on the distortion that could be obtained by any deterministic social choice mechanism, and extend our results on median agent cost to more general objective functions.

This paper presents the underlying theory and the performance of a cluster using a new 2-hop network topology. This topology is constructed using a symmetric equation and Singer Difference Sets and is called SymSig. The degree of connections at each node with SymSig is about half compared to previous methods using Singer Difference Sets. A comparison with a cluster of Clos topology shows significant advantages. The worst case congestion in SymSig topology for unicast permutation is 2, where as in Clos it is proportional to the radix of the building block switches used. The number of switches required is smaller by about 25%, the size of the cluster is larger by about 15% and the worst bandwidth is better by about 50% for SymSig. These advantages are retained for peta and exascale systems. Its performance on a set of collectives like exchange-all, shift-all, broadcast-all and all-to-all send/receive shows improvements ranging from 39% to 83%. Its performance on a molecular dynamics application GROMMACS shows improvement of upto 33%. This network is particularly suitable for applications that require global all to all communications. The low latency of this network makes it scaleable and an attractive alternative for building peta and exascale systems.