Searcharxiv⌕ Search

arXiv subjects

Po Hao Chen

Publications and source records attributed to Po Hao Chen.

3 recordsLinked to original sources

TIDE: Reclaiming Lost Performance In On-Policy Reinforcement Learning

Synchronous, colocated RL post-training of large language models (LLMs) today suffers from severe under-utilization due to long-tail rollouts. Inference engines fall idle waiting for long-tail rollouts or decode at low, inefficient batch sizes. To address this, we present TIDE, an RL training system that preserves on-policy training semantics while recovering resource utilization. TIDE does so by two complementary techniques: (1) dynamically repurposing rollout GPUs for training while long-tail rollouts complete and (2) dynamically migrating rollouts between inference engines to boost decoding throughput. TIDE automatically determines the right time to repurpose rollout resources through an adaptive migration policy, obtaining performance improvements that generalize across multiple applications and hardware configurations without requiring bespoke configuration. In our evaluation, TIDE provides end-to-end speedups of up to 1.36x over colocated systems, outperforming all evaluated on-policy baselines and achieving up to 97% of optimal throughput.

cs.LG↗

A Survey of Potential MPI Complex Collectives: Large-Scale Mining and Analysis of HPC Applications

Offload of MPI collectives to network devices, e.g., NICs and switches, is being implemented as an effective mechanism to improve application performance by reducing inter- and intra-node communication and bypassing MPI software layers. Given the rich deployment of accelerators and programmable NICs/switches in data centers, we posit that there is an opportunity to further improve performance by extending this idea (of in-network collective processing) to a new class of more complex collectives. The most basic type of complex collective is the fusion of existing collectives. In previous work we have demonstrated the efficacy of this additional hardware and software support and shown that it can substantially improve the performance of certain applications. In this work we extend this approach. We seek to characterize a large number of MPI applications to determine overall applicability, both breadth and type, and so provide insight for hardware designers and MPI developers about future offload possibilities. Besides increasing the scope of prior surveys to include finding (potential) new MPI constructs, we also tap into new methods to extend the survey process. Prior surveys on MPI usage considered lists of applications constructed based on application developers' knowledge. The approach taken in this paper, however, is based on an automated mining of a large collection of code sources. More specifically, the mining is accomplished by GitHub REST APIs. We use a database management system to store the results and to answer queries. Another advantage is that this approach provides support for a more complex analysis of MPI usage, which is accomplished by user queries.

cs.DC↗

Big Memory Servers and Modern Approaches to Disk-Based Computation

The Big Memory solution is a new computing paradigm facilitated by commodity server platforms that are available today. It exposes a large RAM subsystem to the Operating System and therefore affords application programmers a number of previously unavailable options for data management. Additionally, certain vendor-specific solutions offer additional memory management options that will result in better data reliability and access speeds.

cs.DC↗