arXiv subjects

Rohan Gandhi

Publications and source records attributed to Rohan Gandhi.

10 recordsLinked to original sources

arXiv2026

CWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms

AI power demand is growing at an unprecedented rate while power grids are often ailing and struggle to keep up. Grid expansion comes with high capital expenditure and long-distance transmission losses, yet there is abundant renewable energy at the source, just not matched to demand. This paper proposes a complementary AI infrastructure deployment model, AI Greeninferencing, that brings modular AI compute to renewable energy sources, focusing on wind, allowing AI footprint expansion, generating local behind-the-meter demand for renewable sites, and helping ease the growing strain on power utilities. Our feasibility analysis shows that 890+ GW of wind capacity lies within 50 ms network round trip time of Azure data centers, and that site-wise right-sizing combined with spatial complementarity of wind energy keeps aggregate fleet utilization on par with traditional deployments. To serve inference requests under variable wind power, we build CWind, a lightweight, reactive, and workload-agnostic AI inference router that uses only real-time signals: inference latency, KV-cache utilization, and queue depth, to dynamically configure sites and distribute requests. Evaluated on a real 64-GPU A100 testbed emulating three wind-powered sites with Azure production traces, CWind reduces P99 end-to-end latency by up to 52% over the strongest contender (also our idea) and by up to 98% over baselines such as power-capping and GPU idling, with consistent gains across workload types, load levels, and GPU generations.

cs.DC↗

arXiv2025

Exposing Weak Links in Multi-Agent Systems under Adversarial Prompting

LLM-based agents are increasingly deployed in multi-agent systems (MAS). As these systems move toward real-world applications, their security becomes paramount. Existing research largely evaluates single-agent security, leaving a critical gap in understanding the vulnerabilities introduced by multi-agent design. However, existing systems fall short due to lack of unified frameworks and metrics focusing on unique rejection modes in MAS. We present SafeAgents, a unified and extensible framework for fine-grained security assessment of MAS. SafeAgents systematically exposes how design choices such as plan construction strategies, inter-agent context sharing, and fallback behaviors affect susceptibility to adversarial prompting. We introduce Dharma, a diagnostic measure that helps identify weak links within multi-agent pipelines. Using SafeAgents, we conduct a comprehensive study across five widely adopted multi-agent architectures (centralized, decentralized, and hybrid variants) on four datasets spanning web tasks, tool use, and code generation. Our findings reveal that common design patterns carry significant vulnerabilities. For example, centralized systems that delegate only atomic instructions to sub-agents obscure harmful objectives, reducing robustness. Our results highlight the need for security-aware design in MAS. Link to code is https://github.com/microsoft/SafeAgents

cs.MA↗

arXiv2025

Improving training time and GPU utilization in geo-distributed language model training

The widespread adoption of language models (LMs) has caused a huge surge in demand for GPUs. Training large LMs requires tens of thousands of GPUs and housing them in the same datacenter (DC) is a challenge due to many constraints including availability of peak power. We focus on training such models across multiple DCs connected via the Wide-Area-Network (WAN). We built Atlas that speeds up the training time using novel workload-aware temporal bandwidth sharing and other design choices. While Atlas improves the training time, it does not completely eliminate the bubbles (idle GPU cycles). We built BubbleTea that runs prefill-as-a-service (part of LM inference) during the bubbles thus improving the GPU utilization without any impact on training. Compared to state-of-the-art designs, Atlas and BubbleTea together achieve up to 17x faster training, and up to 94% GPU utilization. The code will be open-sourced.

cs.DC↗

arXiv2025

BeLLMan: Controlling LLM Congestion

Large language model (LLM) applications are blindfolded to the infrastructure underneath and generate tokens autoregressively, indifferent to the system load, thus risking inferencing latency inflation and poor user experience. Our first-cut controller, named beLLMan, enables the LLM infrastructure to actively and progressively signal the first-party LLM application to adjust the output length in response to changing system load. On a real testbed with H100 GPUs, beLLMan helps keep inferencing latency under control (upto 8X lower end-to-end latency) and reduces energy consumption by 25% (while serving 19% more requests) during periods of congestion for a summarization workload.

cs.DC↗

arXiv2025

Tetris: Efficient Intra-Datacenter Calls Packing for Large Conferencing Services

Conference services like Zoom, Microsoft Teams, and Google Meet facilitate millions of daily calls, yet ensuring high performance at low costs remains a significant challenge. This paper revisits the problem of packing calls across Media Processor (MP) servers that host the calls within individual datacenters (DCs). We show that the algorithm used in Teams -- a large scale conferencing service as well as other state-of-art algorithms are prone to placing calls resulting in some of the MPs becoming hot (high CPU utilization) that leads to degraded performance and/or elevated hosting costs. The problem arises from disregarding the variability in CPU usage among calls, influenced by differences in participant numbers and media types (audio/video), compounded by bursty call arrivals. To tackle this, we propose Tetris, a multi-step framework which (a) optimizes initial call assignments by leveraging historical data and (b) periodically migrates calls from hot MPs using linear optimization, aiming to minimize hot MP usage. Evaluation based on a 24-hour trace of over 10 million calls in one DC shows that Tetris reduces participant numbers on hot MPs by at least 2.5X.

cs.DC↗

arXiv2025

AI Greenferencing: Routing AI Inferencing to Green Modular Data Centers with Heron

AI power demand is growing unprecedentedly thanks to the high power density of AI compute and the emerging inferencing workload. On the supply side, abundant wind power is waiting for grid access in interconnection queues. In this light, this paper argues bringing AI workload to modular compute clusters co-located in wind farms. Our deployment right-sizing strategy makes it economically viable to deploy more than 6 million high-end GPUs today that could consume cheap, green power at its source. We built Heron, a cross-site software router, that could efficiently leverage the complementarity of power generation across wind farms by routing AI inferencing workload around power drops. Using 1-week ofcoding and conversation production traces from Azure and (real) variable wind power traces, we show how Heron improves aggregate goodput of AI compute by up to 80% compared to the state-of-the-art.

cs.DC↗

arXiv2024

Saving Private WAN: Using Internet Paths to Offload WAN Traffic in Conferencing Services

Large-scale video conferencing services incur significant network cost while serving surging global demands. Our work systematically explores the opportunity to offload a fraction of this traffic to the Internet, a cheaper routing option offered already by cloud providers, from WAN without drop in application performance. First, with a large-scale latency measurement study with 3.5 million data points per day spanning 241K source cities and 21 data centers across the globe, we demonstrate that Internet paths perform comparable to or better than the private WAN for parts of the world (e.g., Europe and North America). Next, we present Titan, a live (12+ months) production system that carefully moves a fraction of the conferencing traffic to the Internet using the above observation. Finally, we propose Titan-Next, a research prototype that jointly assigns the conferencing server and routing option (Internet or WAN) for individual calls. With 5 weeks of production data, we show Titan-Next reduces the sum of peak bandwidth on WAN links that defines the operational network cost by up to 61% compared to state-of-the-art baselines. We will open-source parts of the measurement data.

cs.NI↗

arXiv2024

Improving Memory Dependence Prediction with Static Analysis

This paper explores the potential of communicating information gained by static analysis from compilers to Out-of-Order (OoO) machines, focusing on the memory dependence predictor (MDP). The MDP enables loads to issue without all in-flight store addresses being known, with minimal memory order violations. We use LLVM to find loads with no dependencies and label them via their opcode. These labelled loads skip making lookups into the MDP, improving prediction accuracy by reducing false dependencies. We communicate this information in a minimally intrusive way, i.e.~without introducing additional hardware costs or instruction bandwidth, providing these improvements without any additional overhead in the CPU. We find that in select cases in Spec2017, a significant number of load instructions can skip interacting with the MDP and lead to a performance gain. These results point to greater possibilities for static analysis as a source of near zero cost performance gains in future CPU designs.

cs.PL↗

arXiv2024

KnapsackLB: Enabling Performance-Aware Layer-4 Load Balancing

Layer-4 load balancer (LB) is a key building block of online services. In this paper, we empower such LBs to adapt to different and dynamic performance of backend instances (DIPs). Our system, KNAPSACKLB, is generic (can work with variety of LBs), does not require agents on DIPs, LBs or clients, and scales to large numbers of DIPs. KNAPSACKLB uses judicious active probes to learn a mapping from LB weights to the response latency of each DIP, and then applies Integer Linear Programming (ILP) to calculate LB weights that optimize latency, using an iterative method to scale the computation to large numbers of DIPs. Using testbed experiments and simulations, we show that KNAPSACKLB load balances traffic as per the performance and cuts average latency by up to 45% compared to existing designs.

cs.DC↗

arXiv2021

Saath: Speeding up CoFlows by Exploiting the Spatial Dimension

Coflow scheduling improves data-intensive application performance by improving their networking performance. State-of-the-art Coflow schedulers in essence approximate the classic online Shortest-Job-First (SJF) scheduling, designed for a single CPU, in a distributed setting, with no coordination among how the flows of a Coflow at individual ports are scheduled, and as a result suffer two performance drawbacks: (1) The flows of a Coflow may suffer the out-of-sync problem -- they may be scheduled at different times and become drifting apart, negatively affecting the Coflow completion time (CCT); (2) FIFO scheduling of flows at each port bears no notion of SJF, leading to suboptimal CCT. We propose SAATH, an online Coflow scheduler that overcomes the above drawbacks by explicitly exploiting the spatial dimension of Coflows. In SAATH, the global scheduler schedules the flows of a Coflow using an all-or-none policy which mitigates the out-of-sync problem. To order the Coflows within each queue, SAATH resorts to a Least-Contention-First (LCoF) policy which we show extends the gist of SJF to the spatial dimension, complemented with starvation freedom. Our evaluation using an Azure testbed and simulations of two production cluster traces show that compared to Aalo, SAATH reduces the CCT in median (P90) cases by 1.53x (4.5x) and 1.42x (37x), respectively.

cs.DC↗