Pouya Hamadanian

Pouya Hamadanian, Arash Nasr-Esfahany, Malte Schwarzkopf et al.

We study online reinforcement learning (RL) in non-stationary environments, where a time-varying exogenous context process affects the environment dynamics. Online RL is challenging in such environments due to "catastrophic forgetting" (CF). The agent tends to forget prior knowledge as it trains on new experiences. Prior approaches to mitigate this issue assume task labels (which are often not available in practice), employ brittle regularization heuristics, or use off-policy methods that suffer from instability and poor performance. We present Locally Constrained Policy Optimization (LCPO), an online RL approach that combats CF by anchoring policy outputs on old experiences while optimizing the return on current experiences. To perform this anchoring, LCPO locally constrains policy optimization using samples from experiences that lie outside of the current context distribution. We evaluate LCPO in Mujoco, classic control and computer systems environments with a variety of synthetic and real context traces, and find that it outperforms a variety of baselines in the non-stationary setting, while achieving results on-par with a "prescient" agent trained offline across all context traces. LCPO's source code is available at https://github.com/pouyahmdn/LCPO.

Pouya Hamadanian, Pantea Karimi, Arash Nasr-Esfahany et al.

Can an AI autonomously design mechanisms for computer systems on par with the creativity and reasoning of human experts? We present Glia, an AI architecture for networked systems design that uses large language models (LLMs) in a human-inspired, multi-agent workflow. Each agent specializes in reasoning, experimentation, and analysis, collaborating through an evaluation framework that grounds abstract reasoning in empirical feedback. Unlike prior ML-for-systems methods that optimize black-box policies, Glia generates interpretable designs and exposes its reasoning process. When applied to a distributed GPU cluster for LLM inference, it produces new algorithms for request routing, scheduling, and auto-scaling that perform at human-expert levels in significantly less time, while yielding novel insights into workload behavior. Our results suggest that by combining reasoning LLMs with structured experimentation, an AI can produce creative and understandable designs for complex systems problems.

Pouya Hamadanian, Sadjad Fouladi

We introduce Glinthawk, an architecture for offline Large Language Model (LLM) inference. By leveraging a two-tiered structure, Glinthawk optimizes the utilization of the high-end accelerators ("Tier 1") by offloading the attention mechanism to lower-end compute tier ("Tier 2"). This separation allows the memory demand of the attention, known as the key-value cache, to scale independently from the model weights, enabling larger batch sizes and more efficient accelerator usage. Prototyped with NVIDIA T4 GPUs and standard CPU VMs, Glinthawk improves throughput by $5.9\times$ and reduces cost of generation by $2.8\times$, compared to paged attention baselines. For long sequence lengths, it achieves $16.3\times$ throughput improvement at $2.4\times$ less cost. Our evaluation shows that this architecture can tolerate moderate network latency with minimal performance degradation, making it highly effective for latency-tolerant, throughput-focused applications such as batch processing. The prototype is publicly available at https://github.com/microsoft/glinthawk.

Pouya Hamadanian, Malte Schwarzkopf, Siddartha Sen et al.

Recent research has turned to Reinforcement Learning (RL) to solve challenging decision problems, as an alternative to hand-tuned heuristics. RL can learn good policies without the need for modeling the environment's dynamics. Despite this promise, RL remains an impractical solution for many real-world systems problems. A particularly challenging case occurs when the environment changes over time, i.e. it exhibits non-stationarity. In this work, we characterize the challenges introduced by non-stationarity, shed light on the range of approaches to them and develop a robust framework for addressing them to train RL agents in live systems. Such agents must explore and learn new environments, without hurting the system's performance, and remember them over time. To this end, our framework (i) identifies different environments encountered by the live system, (ii) triggers exploration when necessary, (iii) takes precautions to retain knowledge from prior environments, and (iv) employs safeguards to protect the system's performance when the RL agent makes mistakes. We apply our framework to two systems problems, straggler mitigation and adaptive video streaming, and evaluate it against a variety of alternative approaches using real-world and synthetic data. We show that all components of the framework are necessary to cope with non-stationarity and provide guidance on alternative design choices for each component.

Abdullah Alomar, Pouya Hamadanian, Arash Nasr-Esfahany et al.

We present CausalSim, a causal framework for unbiased trace-driven simulation. Current trace-driven simulators assume that the interventions being simulated (e.g., a new algorithm) would not affect the validity of the traces. However, real-world traces are often biased by the choices algorithms make during trace collection, and hence replaying traces under an intervention may lead to incorrect results. CausalSim addresses this challenge by learning a causal model of the system dynamics and latent factors capturing the underlying system conditions during trace collection. It learns these models using an initial randomized control trial (RCT) under a fixed set of algorithms, and then applies them to remove biases from trace data when simulating new algorithms. Key to CausalSim is mapping unbiased trace-driven simulation to a tensor completion problem with extremely sparse observations. By exploiting a basic distributional invariance property present in RCT data, CausalSim enables a novel tensor completion method despite the sparsity of observations. Our extensive evaluation of CausalSim on both real and synthetic datasets, including more than ten months of real data from the Puffer video streaming system shows it improves simulation accuracy, reducing errors by 53% and 61% on average compared to expert-designed and supervised learning baselines. Moreover, CausalSim provides markedly different insights about ABR algorithms compared to the biased baseline simulator, which we validate with a real deployment.

Mehrdad Khani, Pouya Hamadanian, Arash Nasr-Esfahany et al.

Real-time video inference on edge devices like mobile phones and drones is challenging due to the high computation cost of Deep Neural Networks. We present Adaptive Model Streaming (AMS), a new approach to improving performance of efficient lightweight models for video inference on edge devices. AMS uses a remote server to continually train and adapt a small model running on the edge device, boosting its performance on the live video using online knowledge distillation from a large, state-of-the-art model. We discuss the challenges of over-the-network model adaptation for video inference, and present several techniques to reduce communication cost of this approach: avoiding excessive overfitting, updating a small fraction of important model parameters, and adaptive sampling of training frames at edge devices. On the task of video semantic segmentation, our experimental results show 0.4--17.8 percent mean Intersection-over-Union improvement compared to a pre-trained model across several video datasets. Our prototype can perform video segmentation at 30 frames-per-second with 40 milliseconds camera-to-label latency on a Samsung Galaxy S10+ mobile phone, using less than 300 Kbps uplink and downlink bandwidth on the device.