Engineering topics

Pedro, Kevin

Publications and source records attributed to Pedro, Kevin.

4 records

DOE OSTI2024

Using containers to speed up development, to run integration tests and to teach about distributed systems

GlideinWMS is a workload manager provisioning resources for many experiments including CMS and DUNE. The software is distributed both as native packages and specialized production containers. Following an approach used in other communities like web development we built our workspaces, system-like containers to ease development and testing. Developers can change the source tree or check out a different branch and quickly reconfigure the services to see the effect of their changes. In this paper, we’ll talk about what differentiates workspaces from other containers. We’ll describe our base system composed of three containers. A one-node cluster including a compute element and a batch system. A GlideinWMS Factory controlling pilot jobs. And a scheduler and Frontend, to submit jobs and provision resources. Additional containers can be used for optional components. This system can easily run on a laptop and we’ll share our evaluation of different container runtimes, with an eye for ease of use and performance. Finally, we’ll talk about our experience as developers and with students. The GlideinWMS workspaces are easily integrated with IDEs like VS Code, simplifying debugging and allowing development and testing of the system also when offline. They simplified the training and onboarding of new team members and Summer interns. And they were useful in workshops where students could have first-hand experience with the mechanisms and components that, in production, run millions of jobs.

Mambelli, Marco↗

DOE OSTI2024

Simulating the CMS High Granularity Calorimeter with ML

Detector simulation is a key component of physics analysis and related activities in CMS. In the upcoming High Luminosity LHC era, simulation will be required to use a smaller fraction of computing in order to satisfy resource constraints. At the same time, CMS will be upgraded with the new High Granularity Calorimeter (HGCal), which requires significantly more resources to simulate than the existing CMS calorimeters. This computing challenge motivates the use of generative machine learning models as surrogates to replace full physics-based simulation. We study the application of state-of-the-art diffusion models to simulate particle showers in the CMS HGCal. We will discuss methods to overcome the challenges posed by the high-dimensional, irregular geometry of the HGCal. The quality of the showers produced by the diffusion model will be assessed by comparison to the full GEANT4-based simulation. The increase in simulation throughput will be quantified and methods to accelerate the diffusion model inference will also be discussed.

Amram, Oz↗

DOE OSTI2024

Refining Fast Simulation Using Machine Learning

A growing reliance on the fast Monte Carlo (FastSim) will accompany the high luminosity and detector granularity expected in Phase 2. FastSim is roughly 10 times faster than equivalent GEANT4-based full simulation (FullSim). However, reduced accuracy of the FastSim affects some analysis variables and collections. To improve its accuracy, FastSim is refined using regression-based neural networks trained with ML. The status of FastSim refinement is presented. The results show improved agreement with the FullSim output and an improvement in correlations among output observables and external parameters.

Güngördü, Acelya Deniz↗

DOE OSTI2024

Deployment of inference as a service at the US CMS Tier-2 data centers

Coprocessors, especially GPUs, will be a vital ingredient of data production workflows at the HL-LHC. At CMS, the GPU-as-a-service approach for production workflows is implemented by the SONIC project (Services for Optimized Network Inference on Coprocessors). SONIC provides a mechanism for outsourcing computationally demanding algorithms, such as neural network inference, to remote servers, where requests from multiple clients are intelligently distributed across multiple GPUs by a load-balancing service. This talk highlights the recent progress in deploying SONIC at selected U.S. CMS Tier-2 data centers. Using realistic CMS Run3 data processing workflows, such as those containing transformer-based algorithms, we demonstrate how SONIC is integrated into the production-like environment to enable accelerated inference offloading. We will present developments from both the client and server sides, including production job and data center configurations for NVIDIA and AMD GPUs. We will also present performance scaling benchmarks and discuss the challenges of operating SONIC in CMS production, such as server discovery, GPU saturation, fallback server logic, etc.

Holzman, Burt↗