Why Short Prompts Worked While Coding Agents Hung: Fixing GLM-5.2's Queue Lockup

31 July 2026 - 9:00pm

From July 20–22, 2026, our GLM-5.2 setup became unresponsive to incoming coding agent requests under heavy user loads. In this post, we break down why our initial architecture failed under large-context batch workloads and detail the three-part solution—spanning network routing, usage-based priority scheduling, and engine-level prefill tuning—that restored real-time responsiveness across our cluster.

Comparison of Four PDF-Processing Methods

31 July 2026 - 4:30pm

For this blog post, we compared four available PDF-processing methods—pdftotext, docling, scrp-vision, and scrp-marker—on the same 52-page economics research article. The test document is a born-digital manuscript containing body text, multi-column regression tables, display equations, and embedded raster charts. The outputs were evaluated on speed, quality of the text, quality of the tables, quality of the display equations, and quality of the charts.

SCRP's Second-Generation All-Flash Distributed Storage System

8 September 2023 - 8:00pm

Introducing scrp-data, our second generation all-flash distributed storage system. This new, custom-built system will provide four times the capacity and double the sustained throughout of our previous distributed file system, all contained within a 2U rackmount server. Once it is online, this new system will be one of the fastest installed among all academic HPC systems in Hong Kong.

Data Loss on scrp-data: Preliminary Incident Report

20 June 2023 - 10:00am

During the June 2023 maintenance,the cluster has suffered data loss on the distributed storage (/data and ~/large-data) due to software issues. This preliminary report detail what had happened and what we plan to do to avoid similar incidents in the future.

NVIDIA A100 Tensor Core GPU

3 December 2022 - 1:00am

The video version of this post is avaiable here.

Running Multiple RStudio environments on Jupyter

1 September 2022 - 2:00pm

Running RStudio within Jupyter has been possible for quite some time with jupyter-server-proxy. Doing so has its benefits, notably the ability to leverage JupyterHub’s systemdspawner to control the amount of resources users can use, a feature that is not available in the free version of RStudio.

AMD BLAS/LAPACK Optimization in 2022

7 February 2022 - 3:47pm

This is the first of a multi-part series on how to optimize linear algebra performance as of January 2022. We will go over how to optimize BLAS/LAPACK performance on AMD CPUs, focusing on R because it is most affected by this process. In Part 2, we will compare the performance of various software commonly used in economics.