Search for: All records

SLOWPOKE is a new system to accurately quantify the effects of hypothetical optimizations on end-to-end throughput for microservice applications, without relying on tracing or a priori knowledge of the call graph. Microservice operators can use SLOWPOKE to ask what-if performance analysis questions of the form "What throughput could my retail application sustain if I optimized the shopping cart service from 10K req/s to 20K req/s?". Given a target service and its hypothetical optimization, SLOWPOKE employs a perfor- mance model that determines how to selectively slow down non-target services to preserve the relative effect of the optimization. It then performs profiling experiments to predict the end-to-end throughput, as if the optimization had been implemented. Applied to four real-world microservice applications, SLOWPOKE accurately quantifies optimization effects with a root mean squared error of only 2.07%. It is also effective in more complex scenarios, e.g., predicting throughput after scaling optimizations or when bottlenecks arise from mutex contention. Evaluated in large-scale deployments of 45 nodes and 108 synthetic benchmarks, SLOWPOKE further demonstrates its scalability and coverage of a wide range of microservice characteristics. 

Generative recommendation (GR) is an emerging paradigm that tokenizes items into discrete tokens and learns to autoregressively generate the next tokens as predictions. While this token-generation paradigm is expected to surpass traditional transductive methods, potentially generating new items directly based on semantics, we empirically show that GR models predominantly generate items seen during training and struggle to recommend unseen items. In this paper, we propose SpecGR, a plug-and-play framework that enables GR models to recommend new items in an inductive setting. SpecGR uses a drafter model with inductive capability to propose candidate items, which may include both existing items and new items. The GR model then acts as a verifier, accepting or rejecting candidates while retaining its strong ranking capabilities. We further introduce the guided re-drafting technique to make the proposed candidates more aligned with the outputs of generative recommendation models, improving verification efficiency. We consider two variants for drafting: (1) using an auxiliary drafter model for better flexibility, or (2) leveraging the GR model’s own encoder for parameterefficient self-drafting. Extensive experiments on three realworld datasets demonstrate that SpecGR exhibits both strong inductive recommendation ability and the best overall performance among the compared methods. 

Generative recommendation (GR) is an emerging paradigm that tokenizes items into discrete tokens and learns to autoregressively generate the next tokens as predictions. While this token-generation paradigm is expected to surpass traditional transductive methods, potentially generating new items directly based on semantics, we empirically show that GR models predominantly generate items seen during training and struggle to recommend unseen items. In this paper, we propose SpecGR, a plug-and-play framework that enables GR models to recommend new items in an inductive setting. SpecGR uses a drafter model with inductive capability to propose candidate items, which may include both existing items and new items. The GR model then acts as a verifier, accepting or rejecting candidates while retaining its strong ranking capabilities. We further introduce the guided re-drafting technique to make the proposed candidates more aligned with the outputs of generative recommendation models, improving verification efficiency. We consider two variants for drafting: (1) using an auxiliary drafter model for better flexibility, or (2) leveraging the GR model’s own encoder for parameterefficient self-drafting. Extensive experiments on three realworld datasets demonstrate that SpecGR exhibits both strong inductive recommendation ability and the best overall performance among the compared methods. 

Paper in revise and resubmit 

Many-body interactions are essential for understanding non-linear optics and ultrafast spectroscopy of materials. Recent first principles approaches based on nonequilibrium Green’s function formalisms, such as the time-dependent adiabatic GW (TD-aGW) approach, can predict nonequilibrium dynamics of excited states including electron-hole interactions. However, the high-dimensionality of the electron-hole kernel poses significant computational challenges. Here, we develop a data-driven low-rank approximation for the electron-hole kernel, leveraging localized excitonic effects in the Hilbert space of crystalline systems to achieve significant data compression through singular value decomposition (SVD). We show that the subspace of non-zero singular values remains small even as the k-grid grows, ensuring computational tractability with extremely dense k-grids. This low-rank property enables at least 95% data compression and an order-of-magnitude speedup of TD-aGW calculations. Our approach avoids intensive training processes and eliminates time-accumulated errors, seen in previous approaches, providing a general framework for high-throughput, nonequilibrium simulation of light-driven dynamics in materials. 

Claudin-15 (CLDN15) molecules form channels that directly regulate cation and water transport. In the gastrointestinal tract, this transport indirectly impacts nutrient absorption. However, the mechanisms governing ion transport through these channels remain poorly understood. We addressed this question by building on our previous cell culture studies and all atom molecular dynamic simulation model of CLDN15. By mutating D55 to a bulkier glutamic acid (E) or neutral amino acid asparagine (N), our in vitro measurements showed that the D55E mutation decreased charge selectivity and favored small ion permeability, while the D55N mutation led to reduced charge selectivity without markedly altering size selectivity. By establishing a simplified (reduced) CLDN15 molecular dynamics model that excludes non-essential transmembrane regions, we were able to probe how D55 modified cation dehydration, charge interaction, and permeability. These results provide novel insight into organization of the CLDN15 selectivity filter and suggests that D55 plays a dual role in shaping both electrostatic and steric properties of the pore, but its electrostatic role is more prominent in determining CLDN15 cation permeability. This knowledge can be used toward the development of effective strategies to modulate CLDN15 function. The experimental approach established can be further extended to study the function of other claudin channels. Together, these advancements will help us to modulate tight junctions to promote human health. 

We demonstrate efficient spin transfer across a disordered interfacial layer that forms in low damping ferrimagnetic insulator lithium aluminum ferrite (LAFO) and tantalum bilayers. Despite the interfacial disorder, confirmed by transmission electron microscopy, we find a room temperature interfacial spin mixing conductance on the order of 1014 Ω−1m−2 similar to other LAFO-based bilayers with epitaxial interfaces. Broadband ferromagnetic resonance measurements confirm a linewidth broadening in LAFO following the addition of a Ta layer, consistent with the effects of spin pumping. Furthermore, the presence of spin current generated in the Ta layer by spin pumping is confirmed with inverse spin Hall effect measurements. Measurements of the Ta thickness dependence of the spin Hall magnetoresistance and the Gilbert damping enhancement indicate that the Ta spin diffusion length is on the order of 1 nm. This work not only provides a surprising example of efficient spin transport across a disordered interface but also demonstrates the potential for low damping spinel ferrites as a robust system for efficient spin wave spintronics.