Throughput Computing Week 2026

Jun 9, 2026, 12:05 AM → Jun 12, 2026, 3:20 PM America/Chicago

Howard Auditorium (Fluno Center on the University of Wisconsin-Madison Campus)

For the fourth year in a row, Throughput Computing Week 2026 (HTC26) will bring together the Throughput Computing community to share challenges, exchange recent advances and explore opportunities. This year we will welcome the new OSG Executive Director - Peter Couvares - who brings a fresh perspective on how physics communities engage with throughput computing, starting with Multi Messenger Astrophysics.

The OSG Consortium, the HTCondor team, and the Center for High Throughput Computing (CHTC) will be hosting HTC26 at the University of Wisconsin-Madison's Fluno Center. 

This will primarily be an in-person event, but remote participation is also offered. Most speakers will be presenting in person. 

Note: This year, unlike in previous years, HTC26 will begin on a Tuesday (June 9th).

Registration

Registration is now open and is required for attendees, even if you plan to attend remotely only. Registration for in-person attendance will cost $130 per day; there is no fee for registration for virtual attendance. There are two places to register, depending upon whether you will be attending in-person or remotely: 

• Click on the "Register for In-Person Attendance" button on the left in-person attendance registration.
  
  

• Click on the "Register for Virtual Participation" button on the left for remote attendance registration.
  
  

 Registration for in-person attendance closes May 31. 

In-Person Attendance

In-person registration is $130 per day and includes a full lunch buffet and a light breakfast, snacks and coffee and drinks at the breaks. With in-person registration, you can attend plenary and non-plenary sessions, mingle with colleagues, and have planned or ad hoc meetings. 

In addition, recreational and evening events are planned throughout the week, including a social gathering on the Memorial Union Terrace. 

Remote Attendance

Plenary presentations (only) will be broadcast via Zoom. There is no registration fee for remote attendance. 

Event Details

HTC26 gives OSPool, OSDF, HTCondor and Pelican users, contributors, and collaborators the chance to exchange ideas and experiences, to learn about the latest services and research, and to experience live demos. Current or potential consumers or providers of distributed high throughput computing, HTCondor, OSDF, or Pelican users or administrators are welcome to attend.

This year HTC26 will include a focus on a topics including:

• The challenges of AI: 
  • Successes and Failures in using Agentic AI
  • AI Facilitation 
  • Best practices for AI driven Research
  • Using AI tools to understand and debug HTCondor and Pelican

• The Challenges of Multi Messenger Astrophyics Computing
• Facilitation for HTC
• HTC Admin Jumpstart and Masterclass
• Throughput Computing at Work: Reports from the Field
• The Challenges of Data and the Challenges of Architecture (i.e., GPU's, Arc and Memory)

 

In additon, following the success of the last two years, CC* (current or future) campuses and administrators can attend dedicated break out sessions on Wednesday, June 10 of the event.

HTC26 will include presentations from researchers and administrators who use OSG services,  HTCondor, OSDF or Pelican. Leadership will discuss the current state and future plans for these services. Tutorials across the services will be offered. 

Hotel Reservations

We have hotel blocks reserved at nearby hotels, as well as at the conference venue. Please do not delay booking your room as there are a limited number that are at a reduced rate. Click on the "Local Arrangements" button on the left for details.

Please consider speaking!

We encourage attendees to share their thoughts and experiences by signing up to be a speaker; click on the "Speaker Information" button on the left for more information. We enjoy hearing from members of the community!

Consultations

Need help with a specific HTCondor or OSPool issue? 

Set up a one-on-one consultation with an HTCondor developer at HTC26! To do so, please email the HTCondor team (htcondor-week@cs.wisc.edu) with some information about your issue, and we'll match you up with an appropriate HTCondor person. Similarly, to set up a consultation with an OSPool representative or facilitator, please email support@osg-htc.org. To set up a consultation with a Pelican representative, please email pi-team@pelicanplatform.org. 

 

 

Tue, June 9

6:30 AM → 7:30 AM Lakeshore Run

7:30 AM → 9:00 AM Registration, Coffee, and Light Breakfast

9:00 AM → 10:30 AM The PATh Forward

9:00 AM Welcome to Throughput Computing Week

Speaker: Dr Patrick McDaniel (UW-Madison)

9:05 AM To Be Provided

Speaker: Dr Miron Livny (UW-Center for High Throughput Computing)

9:35 AM A Perspective from Smaller Institutions

Speaker: Jason Simms (Swarthmore College)

10:05 AM TBD; Subject CI Compass

Speaker: Ewa Deelman (USC Information Sciences Institute)

9:01 AM → 10:30 AM US ATLAS Meeting

10:30 AM → 11:00 AM Break

11:00 AM → 12:30 PM Facilitation for HTC, AI Challenges and HTC AI Toolkit, David Swanson Awardee Researcher Talk

11:00 AM LLM Assisted Facilitation

Speaker: Christina Koch (UW Madison)

11:25 AM YOL-No: Guiding Agent Usage on the Access point

Speaker: Ian Ross (U. Wisconsin)

11:50 AM Introduction of David Swanson Awardee

Speaker: Ronda Swanson (University of Nebraska-Lincoln)

12:00 PM David Swanson Award and Researcher Presentation

Speaker: Saloni Bhogale (University of Wisconsin-Madison)

12:30 PM → 1:30 PM Lunch

1:30 PM → 2:35 PM The Promise of Multi-Messenger Astronomy: Transforming Research (Keynote)

1:30 PM Keynote Introduction

1:35 PM Keynote: The Present and the Future of Multi-Messenger Astrophysics

With the detection of compact binary coalescences and their electromagnetic counterparts by gravitational-wave detectors, a new era of multi-messenger astronomy has begun. In this talk, I will describe how data science is enabling the gravitational-wave community to make very low-latency detection and parameter estimation possible within the alert system. I will then discuss how current ground based optical surveys and dedicated follow-up systems are integrating machine learning into their standard work flows, with examples from both current and near future surveys. We will close with near-term prospects for the field.

Speaker: Dr Michael Coughlin (University of Minnesota)

2:35 PM → 3:05 PM Break

3:05 PM → 5:00 PM Multi-Messenger Astrophysics Computing Challenges

3:05 PM Title - TBD: Topic - Real-Time Pipeline Latency

Speaker: Kendall Ackley (University of Warwick/Gravitational-wave Optical Transient Observer)

3:30 PM Title: TBD Topic: ZTH/Rubin Kilonova Searches (Remote Presentation)

Speaker: Shreya Anand (Standford University)

3:50 PM On the role of multimodality for the real-time discovery and classification of transients

The Rubin Observatory is detecting an unprecedented number of transients each night, exceeding the classification capacity of all existing spectroscopic resources my several orders of magnitude. As a result, robust photometric classifications will be essential both for assembling complete samples of different transient subtypes and for identifying events that warrant spectroscopic follow-up.

In this talk, we introduce a family of highly performant, real-time hierarchical classifiers designed for ZTF and LSST alert streams. These models incorporate images, multi-band light curves, and contextual metadata, to deliver reliable, high-level classifications within seconds of the first alert. Our hierarchical approach is unique in its ability to make classifications at several levels of granularity based on the available information, making it possible, for the first time, to produce high-level classifications at early times, and refine those classifications as more observations become available.

Our models achieve strong performance on real ZTF data, reaching >95% accuracy on binary Transient vs Persistent classification after the first day. For a more granular task distinguishing CVs, AGNs, and supernova subtypes, our models achieve >85% accuracy 128 days after first detection with our multimodal model delivering performance that far exceeds the light-curve only models, especially at early times. On the ELAsTiCC datasets of simulated LSST observations, we achieve state-of-the-art performance with a macro F1 score of 0.88. These results demonstrate that multi-modal, hierarchical classifiers can deliver classifications at LSST scale, supporting both real-time triaging and population studies from the earliest alerts.

Speaker: Ved Shah (Northwestern University)

4:15 PM TBD (Remote Presentation)

Speaker: Gautham Narayan (University of Illinois Urbana-Champaign)

4:35 PM Constraining Degeneracies in Gamma-Ray Burst Simulations

I will talk about difficulties associated with free parameters in modeling GRB emission. Without a way to constrain these parameters, inferring too much information from models can be problematic. Recent advances in the physics that can launch GRB jets, however, have the potential to enable a deep understanding of GRBs and their progenitors. My talk will address how degeneracies arise in simulations, with an emphasis on how they may be resolved.

Speaker: Nathan Walker (Oregon State University)

3:05 PM → 5:00 PM US ATLAS Meeting: US ATLAS Meeting - Tuesday PM

5:00 PM → 5:05 PM Closing Remarks

5:05 PM → 7:35 PM Evening on the Memorial Union Terrace

We will walk over together to the Memorial Union Terrace following the last session for dinner. At the Terrace you can purchase burgers, brats, beer, and other casual fare. The rain date will be the following day - Wednesday, June 10.

Wed, June 10

8:00 AM → 9:00 AM Coffee, Pastries and Registration

9:00 AM → 10:30 AM Multi-Messenger Astrophysics Computing Challenges Part 2

9:00 AM Title TBD; Subject: Rubin and its Computing Needs

Speaker: Colin Slater

9:25 AM Of Needles and Haystacks: Lessons from the GW-MMADS follow-up campaign during the fourth LIGO/Virgo/KAGRA observing run

Gravitational wave follow-up requires iterative observation, analysis, and decision-making on the order of hours in the most extreme of cases. To this extent, it is a subfield where throughput is defined not only by the amount of pertinent data, but rather the speed at which the information in the data can be synthesized into a human-digestible format. In this talk I will present the follow-up effort of the Gravitational Wave-MultiMessenger Astronomy DECam Survey team (PIs Andreoni+Palmese) during the fourth LIGO/Virgo/KAGRA observing run through the lens of considering the technological infrastructure required to perform effective GW follow-up. I propose that maximizing scientific yield of follow-up efforts relies on maintaining a human-centric perspective on pipeline development in which the aggregation and presentation of big data products are just as important as the analytical methods producing said products.

Speaker: Tomas Cabrera (Carnegie Mellon University)

9:50 AM Radio Follow-up of GW Events: Opportunities and Challenges (Remote Presentation)

In this talk, I will discuss current and future opportunities in the radio follow-up of gravitational-wave (GW) events using interferometric, PI-driven facilities such as the Karl G. Jansky Very Large Array (VLA) and the future next-generation VLA (ngVLA). I will highlight lessons learned from recent GW observing runs, and strategies for identifying and characterizing relativistic ejecta and multi-messenger counterparts across a broad range of transients. I will also discuss challenges related to observation planning, data reduction, sensitivity limitations, and long-term monitoring requirements. I will conclude with prospects for multi-messenger astronomy in the ngVLA era.

Speaker: Alessandra Corsi (John Hopkins University)

10:10 AM The VLA Sky Survey and multi-messenger astronomy

The VLA Sky Survey images the whole sky at 2-4GHz in the radio with angular resolution comparable to optical telescopes, providing a unique reference survey for the radio sky. Multi-epoch data allows variable sources and transients to be found and characterized and spectral and polarimetric information is also recorded. Processing the large amount of data from this survey is a significant computational challenge as the high resolution of the survey means that subtle calibration and geometric effects need to be carefully treated. In this talk I will describe both some of the multi-messegenger science enabled by VLASS and the computational and logistical challenges of processing VLASS and other NRAO data in HPC/HTC environments.

Speaker: Mark Lacy

9:00 AM → 10:30 AM US ATLAS Meeting

10:30 AM → 11:00 AM Break

11:00 AM → 12:30 PM Campus Impacts and Broader Outreach

11:00 AM Campus Contributors: What's New?

Speakers: Christina Koch (UW Madison), Tim Cartwright (University of Wisconsin–Madison, OSG)

11:25 AM A Statewide Perspective from a REN

Speaker: Dr Pierrette Dagg (MERIT)

11:50 AM Trends in campus resource planning and resource sharing

As AI demands boom, and the market forces we all face as a side effect continue to impact campus buying power, it is more important than ever to understand how to capacity plan for your campus, and how that capacity planning fits into the larget picture of your state, region, and the nation. This presentation will share research in capacity planning for campuses, and explore how shared computing models continue to be a key way for campuses to provide resources in the AI era.

Speaker: Preston Smith (Purdue University)

12:10 PM Community-Operated CI Resources for Campus Researchers

Speaker: Stephen Deems

11:00 AM → 12:30 PM Throughput Computing at Work: Reports from the Field

11:00 AM The IceCube CredMon - Going Beyond the Manual to Support Pelican

If the built-in HTCondor OAuth2 doesn't meet your needs, you can always choose to do your own thing. In this talk, we show how and why IceCube wrote a custom token storer and CredMon to handle OAuth2 token creation and refresh. While custom code allows exactly matching the IceCube workflow, the main defining feature is not having to ask command line users to do an additional web login. Along the way, we worked around several limitations in HTCondor and noted some things to avoid.

Speaker: David Schultz (University of Wisconsin-Madison)

11:25 AM Rubin Observatory: Scaling towards Data Release Production using HTCondor

Rubin Observatory will begin conducting the Legacy Survey of Space and Time (LSST), a 10 year survey of the Southern sky. This presentation will briefly cover how the project successfully uses HTCondor based workflows for several flavors of processing, including preparations for the project's Data Release Production at the United States Data Facility.
The project's middleware that interfaces between the user and HTCondor,  a.k.a., Batch Production Service (BPS), will be introduced as well as the "allocateNodes" mechanism being used to submit user glideins to Slurm clusters.

Speaker: Greg Daues (NCSA)

11:50 AM JAWS and the Backend Architecture: HTCondor and Pool Management for Portable Scientific Workflows

Authors

Seung-Jin Sul*, Mario Melara, Ramani Kothadia, Ludovico Bianchi, Joshua Boverhof, Nick Tyler, Daniela Cassol, Mike Sneddon, Setareh Sarrafan, Kjiersten Fagnan
Lawrence Berkeley National Laboratory, Berkeley, CA, USA

Abstract

Scientific workflows at the DOE Joint Genome Institute (JGI) increasingly need to run across multiple high-performance computing facilities, but each site exposes a different mix of schedulers, queue policies, container runtimes, data-transfer paths, and security constraints. The JGI Analysis Workflow Service (JAWS) addresses this problem by separating the scientific workflow interface from the facility-specific execution backend. Users submit Workflow Description Language (WDL) workflows through a common client and JAWS Central service, while each JAWS Site provides the local backend configuration needed to execute those workflows on a particular compute environment.

This talk focuses on an introduction to JAWS and the backend architecture that makes that portability practical, with emphasis on HTCondor and the JAWS Pool Manager. In JAWS, Cromwell orchestrates WDL workflows, and HTCondor serves as the execution backend that receives task submissions, maintains the job queue, and dispatches work to compute pools at the target site. The JAWS Pool Manager implements the HTCondor glidein: it monitors HTCondor job queue, determines whether the pool has capacity, and provisions worker nodes from the site's native batch system (Slurm or PBS) to scale the HTCondor pool up and down with demand. The JAWS and HTCondor logs and metadata support performance reporting for CPU use, memory efficiency, runtimes, and failure modes. This design enables JAWS run task-by-task execution over heterogeneous facilities while presenting a consistent operational model to users and workflow developers.

Keywords

JAWS, HTCondor, workflow management, WDL, Cromwell, pool manager, HPC, DOE computing

Acknowledgement

The work conducted by the U.S. Department of Energy Joint Genome Institute, a DOE Office of Science User Facility, is supported by the Office of Science of the U.S. Department of Energy operated under Contract No. DE-AC02-05CH11231.

Speaker: Seung-Jin Sul (Staff Software Engineer, U.S. Department of Energy (DOE) Joint Genome Institute (JGI), Lawrence Berkeley National Laboratory)

12:10 PM Evolving the CMS Submission Infrastructure

The CMS Submission Infrastructure (SI), based on HTCondor and GlideinWMS, operates a federated pool of resources, integrated from WLCG, HPC and cloud providers, supporting CMS offline computing needs. As CMS prepares for the High-Luminosity LHC (HL-LHC) phase, the infrastructure must continue evolving to support increasing resource scale and heterogeinity, while maintaining the stability, flexibility, and efficiency required for large-scale distributed computing.

This contribution presents recent developments in the CMS SI, focused on improving resource utilization and scheduling efficiency. In particular, we discuss the deployment of high-IO auxiliary slots to reduce slot fragmentation caused by bursts of short single-core workloads, and the introduction of controlled pilot overloading techniques aimed at recovering otherwise unused CPU cycles from inefficient or I/O-bound payloads. These strategies have shown systematic improvements in overall CPU efficiency while preserving stable SI operations..

The talk will also cover the continued evolution of CMS resource integration strategies for heterogeneous computing environments, including GPUs, ARM processors, and HPC facilities, together with broader CMS plans for the next-generation Workflow Management system designed for HL-LHC-scale workflows and increasingly dynamic computing resources. Together, these developments represent important steps toward a more scalable, sustainable, and efficient CMS computing infrastructure, as required by the HL-LHC program.

Speaker: Marco Mascheroni (UCSD)

11:00 AM → 12:30 PM US ATLAS Meeting

12:30 PM → 1:30 PM Lunch

1:30 PM → 2:30 PM Sharing Data, Data Stores and Utilizing the OSDF

1:30 PM Booster rockets for Birds: New Backends for the Pelican Origin

Speaker: Justin Hiemstra (Morgridge Institute for Research)

1:55 PM Sharing Data: Authentication, Users, and Group Management in Pelican

Speaker: Brian Aydemir (Morgridge Institute for Research, UW–Madison)

2:15 PM Open Science data federation and National Research platform

The Open Science Data Federation (OSDF) and the National Research Platform (NRP) represent a transformative shift toward a democratized, high-performance computational ecosystem designed to accelerate global scientific discovery. By integrating the OSDF’s robust data distribution capabilities with the NRP’s distributed, Kubernetes-based "Nautilus" infrastructure, researchers can seamlessly orchestrate data-intensive workflows across a federated backbone of compute and networking resources. As a community-driven cyberinfrastructure, the combined OSDF and NRP framework provides a scalable, secure, and vendor-neutral environment that bridges the gap between heterogeneous data repositories and high-speed processing, ultimately fostering a more collaborative and efficient open science landscape.

Speaker: Fabio Andrijauskas

1:30 PM → 2:30 PM US ATLAS Meeting

2:30 PM → 3:00 PM Break

3:00 PM → 4:30 PM Vibe-coding the future: Successes and Dramatic Failures in using Agentic AI

3:00 PM Agents and Birds - Leveraging LLMs for Throughput Computing

Speaker: Brian Bockelman (Morgridge Institute for Research)

3:25 PM Deploying Nextflow on CHTC: A Case Study for Serving Biology's Software Needs with AI-Assisted Engineering

Speaker: Nicholas Minor (Dave O'Connor's Laboratory/UW-Madison)

3:50 PM A Tale of Two Ecosystems: Model Training in the OSPool and NAIRR

Speaker: Ian Ross (U. Wisconsin)

4:10 PM Pegasus & AI Technologies

Speaker: Dr Ewa Deelman (USC Information Sciences Institute)

4:30 PM → 4:45 PM CHTC Fellows Lightning Talks

4:45 PM → 4:55 PM Closing Remarks

5:30 PM → 7:15 PM Small Group Dinner

At the close of the day's sessions, we'll announce small group dinner options to join. A couple CHTC staff members will lead each group to an area restaurant. Options are typically reasonably priced either on campus or the Capital Square.

6:00 PM → 7:15 PM Evening Bike Ride

A relaxed casual ride through the Arboretum. We will depart from the Fluno Center at 6 pm. CHTC staff will be around to assist with e-bike rentals from area stations.

7:30 PM → 9:30 PM Bonfire at Picnic Point with Smores and Hotdogs

We'll have staff available for rides to Lake Shore from Fluno. We'll have the fixins to roast hot dogs and make smores. Please bring your own water bottles or sodas, if desired No alcohol permitted.

Thu, June 11

8:00 AM → 9:00 AM Coffee, Pastries and Registration

9:00 AM → 10:30 AM Keeping an Eagle Eye On Things (monitoring)

9:00 AM Using OSG to enable the development of Simulation based inference for Cosmology

I will describe how OSG has enabled us to develop a simulation-based inference (SBI) method for cosmology analysis. For this application, we train an SBI model, based on a mixture density network (MDN), to derive posteriors for cosmological parameters from a data vector that describe observations of galaxy clusters in the universe. We use analytic models to generate mocks of the observational data vectors needed for building and testing the SBI method, and OSG enabled us to generate these large training and testing data sets.

Speaker: Yuanyuan Zhang (NSF NOIRLab)

9:25 AM OSG Security Status and The Future

This will be a high level overview of the security team's activities on OSG, what we're responsible for, and where we're going. It will provide summary metrics of what we've done, a short overview of the security exercise we've performed for US CMS and US ATLAS, and talk about us preparing for the coming wave of super AI that will discover all the zero days all at once.

Speaker: Mark Krenz (Indiana University)

9:50 AM OrangeGrid Cloud Workers

OrangeGrid Cloud Workers is the infrastructure that lets Syracuse University's on premises HTCondor cluster overflow into public cloud spot instances when campus capacity runs out. A researcher submits a job the same way they always have, condor_submit my.sub from a campus submit host, and if the job opts in with +WantCloudBurst = "...", it can land on an ephemeral VM running in any public or participating private cloud. The cloud VM joins the cluster as a regular HTCondor execute node, mounts the user's NFS home directory, authenticates with campus LDAP, and runs the job. When the spot instance is reclaimed, another worker takes over. No new tooling for users, no data migration, no second workflow.

This solves a specific problem. Campus compute is finite, demand isn't. Cloud bursting can help with this. The hard part is doing it without breaking the user facing contract around homedirs, UIDs, and job submission.

.

Speaker: Peter Pizzimenti

10:30 AM → 11:00 AM Break

11:00 AM → 12:30 PM CHTC Work

11:00 AM From Per-Job Data to an Aggregated Workload Insight: A Toolkit for Profiling HTCondor Workload

Researchers using HTCondor for high-throughput computing routinely submit groups of related jobs, known as Clusters, ranging from hundreds to tens of thousands of jobs each. Current tools report per-job data, making it difficult to diagnose Cluster-wide issues such as jobs stuck on hold, poor resource utilization, or unexpected failures. We present a Python toolkit, to be included as a part of the HTCondor suite, that bridges this gap. Given a single Cluster ID, the toolkit evaluates job status distribution, runtime patterns, hold reasons, and resource utilization, then synthesizes these into a colour-coded health summary that tells researchers and facilitators whether their Cluster is running well, and if not, why. Each analysis transforms thousands of individual job records into a concise, actionable report.

Speaker: Kashika Mahajan (UW-Center for High Throughput Computing)

11:25 AM The HTCon-Door: Opening Your Compute Cluster to the World

Connecting High Throughput batch resources to Open OnDemand for access via a web browser anywhere/anytime.

Speaker: Cole Bollig (UW-Center for High Throughput Computing)

11:50 AM Common Input Files

Speaker: Todd Miller (CHTC)

12:15 PM Authorizing Remote Placement with the PlacementD

Speaker: Matyas Selmeci (UW-Madison CHTC)

12:30 PM → 1:30 PM Lunch

1:30 PM → 3:00 PM Workflows, near and far

1:30 PM GlideinWMS Improvements and Challenges

GlideinWMS is a workload management and resource provisioning system widely used for distributed scientific computing. Recent development efforts have focused on simplifying token-based authentication which is finally starting to be adopted in production and improving support for High Performance Computing (HPC) resources. This presentation will highlight recent progress in these areas and discuss opportunities where additional collaboration with the HTCSS community could provide significant benefits.

A redesigned credential framework and new token generators simplify configuration. For example by enabling credentials and authentication options to be selected dynamically based on job ClassAds and target resources. Building on this improved infrastructure, GlideinWMS Frontend now supports automatic generation of self-signed SciTokens, providing a lightweight solution particularly well suited for testing environments and smaller virtual organizations (VOs).

Expanding access to HPC systems remains essential for maximizing available scientific computing capacity. With support and collaboration from the HTCSS community, GlideinWMS is broadening HPC integration through both established mechanisms, such as remote cluster access, and modern REST-based interfaces including SFApi and IRI Facility API. This work aims to reduce operational complexity while enabling more flexible and scalable access to heterogeneous computing resources.

Speakers: Marco Mambelli (Fermilab), Namratha Urs (Fermi National Accelerator Laboratory)

1:55 PM Auto-regulating input files that save time when jobs fail

I use CHTC and OSPool to run a rainfall frequency analysis pipeline across 200+ watersheds as part of FEMA's National Flood Insurance Program work. DAGMan was essential for automating job submission and output processing at this scale, but managing pipeline failures proved challenging -- specifically, ensuring that every daily precipitation file was properly analyzed before downstream steps proceeded. A failure-and-retry approach led to redundant computation, while simply continuing risked carrying missing or corrupted data forward into later analyses. To address this, I developed submit files that intake dynamically updated file lists, whittling down as outputs are confirmed present. This prevents the pipeline from advancing with incomplete data while also avoiding redundant reprocessing of files that completed successfully. I could also focus instead more on the pipeline itself if that would be preferred. I am available to present in person on 6/10 or 6/11.

Speaker: Benjamin FitzGerald (University of Wisconsin-Madison)

2:15 PM Heterogeneous Prompt Processing Challenges in DUNE

DUNE faces unique challenges for data management and movement due to its very large event record size, needs for immediate prompt processing, and requirements for heterogeneous computing across an internationally distributed compute and storage system. We describe DUNE's current use of high throughput and high performance computing, including the growing AI/ML component, and the challenges we face in getting the data, code, and database information to the right kind of computer in timely fashion.

Speaker: Steven Timm (Fermilab (DUNE))

2:40 PM Wrangling Massive Task Graphs with VineReduce

A number of modern programming frameworks encourage end users to write concurrent functional programs that are expanded into task graphs, and then executed using local parallelism. While providing an elegant user experience, these systems struggle when presented with large programs that generate million-node graphs and must run on heterogeneous systems. We demonstrate a new framework, VineReduce, that allows for the transformation of these programs into an adaptive hierarchy that can effectively harness distributed, heterogeneous HTCondor pools.

Speaker: Douglas Thain (University of Notre Dame)

3:00 PM → 3:30 PM Break

3:30 PM → 4:45 PM HTCSS - the Migratory Path: Where We Have Been, and Where We are Going

3:30 PM HTCondor Flight Path

HTCondor discussion of what's new and what coming up.

Speaker: Todd Tannenbaum (University of Wisconsin)

4:15 PM Todd or Bot Game Show

Speakers: Greg Thain (Center for High Throughput Computing), Todd Tannenbaum (University of Wisconsin)

4:45 PM → 5:00 PM Closing Remarks

5:00 PM → 6:30 PM Small Group Dinners

6:00 PM → 9:00 PM 80's Tribute Night at the Monona Terrace rooftop featuring the band Sixteen Candles

Todd Tannenbaum is leading a group to this event. PLEASE NOTE YOU MUST RESERVE TICKETS. The event is free. Gates & Cash Bars Open at 5:30. Performance Time: 7:00 - 9:00 PM. Tickets can be reserved online via Eventbrite(https://www.eventbrite.com/e/concerts-on-the-rooftop-sixteen-candles-80s-tribute-tickets-1987934688166) or by calling 608-261-4062.

6:00 PM → 7:30 PM Sunset Kayak or Paddleboard

8:45 PM → 10:45 PM Karaoke Night

Aaron Moate will lead a group across the street from Fluno's to Mom's Bar, one of the best Karaoke Bar's in the area. If you are joining late from another event, the Bar is right across the street from Fluno at 614 University Ave.

Fri, June 12

8:00 AM → 9:00 AM Coffee, Pastries, and Registration

9:00 AM → 10:30 AM AI Facilitation, User Challenges, and best practices. + AI driven Research

9:00 AM TBD

Speaker: Jingyan Shi (INSTITUTE OF HIGH ENERGY PHYSICS, Chinese Academy of Science)

9:20 AM Climate Policy Modeling with HTCondor

HTCondor was used to develop DPSOL, a framework for solving dynamic programming problems. It was the foundation for DSICE which merged dynamic and stochastic factors in economics and the climate, and analyzed the social cost of carbon, the cost of a two-degree target and the value of carbon capture and sequestration.

Speaker: Ken Judd (Stanford University)

9:45 AM Where's the Big Data for Protein Dynamics?

Speaker: Hannah Wayment-Steele (UW-Madison)

10:10 AM Enabling Large-Scale and Ultra-Large AlphaFold3 Predictions with Distributed High-Throughput Computing

AlphaFold 3 (AF3) enables atomic-resolution prediction of biomolecular complexes, driving rapidly growing demand across the life sciences. However, its ~750 GB reference database has effectively confined production deployments to systems with shared parallel filesystems, creating a major barrier for scalability. Distributed high-throughput computing (dHTC) platforms offer vast, heterogeneous compute capacity but fundamentally lack the shared data infrastructure assumed by AF3. We present a data-aware deployment of AF3 for dHTC, implemented on the Center for High Throughput Computing (CHTC) and the Open Science Pool (OSPool). The workflow is decomposed into a CPU-bound data pipeline that executes on nodes with locally staged, scheduler-advertised databases, and a GPU-bound inference pipeline that opportunistically scales across distributed resources. Using CUDA Unified Virtual Memory (UVM), we extend inference beyond physical GPU limits, enabling predictions of ultra-large complexes that exceed device VRAM. By elevating dataset locality to a schedulable resource via HTCondor ClassAds, we eliminate prohibitive per-job data transfers and enable efficient, federated execution. Beyond scaling throughput, we demonstrate that dHTC can support previously infeasible workloads. Together, these results establish dHTC as a viable—and in some regimes superior—execution model for data-intensive structural biology workflows and provide a general blueprint for deploying large, data-intensive applications on distributed cyberinfrastructure.

Speaker: Daniel (Danny) Morales (UW-Center for High Throughput Computing)

10:30 AM → 11:00 AM Break

11:00 AM → 12:20 PM Is AI for The Birds? Using AI tools to understand and debug HTCondor and Pelican

11:00 AM MCP for HTCondor Admins

Speaker: Rob Gardner (University of Chicago)

11:25 AM Cluster Metrics for External Services

Discussion of metrics that Condor can provide about the performance of external services. Condor has a unique view of the performance of the services that it uses on behalf of jobs. Examples of external services include file transfer plugins and credmons. Individual failures are not very interesting to cluster administrators, but widespread failures affecting many jobs are. What sort of metrics might be useful?

Speaker: Ron Tapia (Penn State University)

12:20 PM → 12:30 PM Closing Remarks

12:30 PM → 1:30 PM Lunch

1:30 PM → 3:00 PM No Session