<?xml version="1.0" encoding="utf-8"?> <feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en"> <title>Field notes · Anthony Kougkas</title> <subtitle>Field notes by Anthony Kougkas.</subtitle> <id>https://akougkas.io/feed/field-note.xml</id> <link rel="self" type="application/atom+xml" href="https://akougkas.io/feed/field-note.xml"/> <link rel="alternate" type="text/html" href="https://akougkas.io/journal/?kind=field-note"/> <updated>2026-11-08T00:00:00-06:00</updated> <author> <name>Anthony Kougkas</name> <uri>https://akougkas.io/about/</uri> </author> <entry> <title type="html">SC26 in Chicago</title> <link rel="alternate" type="text/html" href="https://akougkas.io/journal/2026/sc26-in-chicago/"/> <id>https://akougkas.io/journal/2026/sc26-in-chicago/</id> <published>2026-11-08T00:00:00-06:00</published> <updated>2026-11-08T00:00:00-06:00</updated> <summary type="html">What the group brings to SC26 in Chicago, the pressure from GPUs and agent fleets that runs through all of it, and an invitation to Thursday&apos;s Birds of a Feather on HDF5 in the age of AI-driven science.</summary> <content type="html" xml:base="https://akougkas.io/journal/2026/sc26-in-chicago/">&lt;p&gt;SC26 runs November 15 to 20 at McCormick Place in Chicago. At the end of my &lt;a href=&quot;/journal/2025/reflecting-on-sc25/&quot;&gt;SC25 note&lt;/a&gt; I wrote down what I wanted from this year: the students in the building, each with something to present. Here is what the group is bringing, and the question I see running through all of it.&lt;/p&gt; &lt;p&gt;Read the abstracts side by side and the same pressure shows up from two directions. GPUs now want to read and write where their data lives, and fleets of LLM agents now read scientific data on their own schedule. The I/O stack was not built for either consumer, and each piece of this work restores something the stack loses when they arrive: a data model, an identity for an operation, a check on whether the answer is still science.&lt;/p&gt; &lt;h2 id=&quot;the-gpu-side&quot;&gt;The GPU side&lt;/h2&gt; &lt;p&gt;On Monday at &lt;a href=&quot;https://sc26.conference-program.com/presentation/?id=ws_pdsw111&amp;amp;sess=sess129&quot;&gt;PDSW 2026&lt;/a&gt;, J. Fernandes and &lt;a href=&quot;https://grc.iit.edu/center/team/luke-logan/&quot;&gt;Luke Logan&lt;/a&gt; present &lt;a href=&quot;/publications/fernandes2026gpuh5/&quot;&gt;GPUH5&lt;/a&gt;. A kernel cannot write where its data lives: it has to exit and copy its array to the host before HDF5 can run. Systems that let a GPU start I/O give up the data model. GPUH5 keeps HDF5’s hierarchical paths and chunked arrays and replaces the container with a key-value store, so an in-kernel write reduces to enqueuing a pointer. On a Gray-Scott simulation writing 380 snapshots to NVMe, it hides 96% of the I/O its synchronous twin exposes and finishes 1.9x faster than default HDF5.&lt;/p&gt; &lt;p&gt;The same morning &lt;a href=&quot;https://grc.iit.edu/center/team/izzet-yildirim/&quot;&gt;Izzet Yildirim&lt;/a&gt; presents &lt;a href=&quot;/publications/yildirim2026gdsight/&quot;&gt;GDSight&lt;/a&gt;. GPUDirect Storage copies NVMe data straight into GPU memory, and in doing so it drops the link between an application operation and the device work it causes. GDSight uses eBPF to restore that link for each NVMe command. On unmodified KvikIO, NIXL, and HDF5 it pins a silent bypass that cuFile’s counters miss, 4x device amplification, and 22x tail inflation to their causes.&lt;/p&gt; &lt;h2 id=&quot;the-agent-side&quot;&gt;The agent side&lt;/h2&gt; &lt;p&gt;Izzet also has two posters on Tuesday, and both start from the same observation: scientific workflows increasingly run as fleets of agents analyzing immutable datasets. &lt;a href=&quot;/publications/yildirim2026stageonce/&quot;&gt;StageOnce&lt;/a&gt; coordinates the agents’ speculative path predictions through a node-local daemon, so concurrent misses become one immutable copy. In replays of genomics and astronomy sessions it cuts fleet makespan by 31 to 51% at 12 concurrent jobs. &lt;a href=&quot;/publications/yildirim2026toolfuse/&quot;&gt;ToolFuse&lt;/a&gt; coalesces equivalent tool calls while they are still in flight, using an execution identity built from versioned dependencies rather than arguments alone; 16 equivalent requests become one execution. Both papers also say plainly when their technique does not help, which is the part I would read first.&lt;/p&gt; &lt;p&gt;The other two posters come from &lt;a href=&quot;https://grc.iit.edu/center/team/neeraj-rajesh/&quot;&gt;Neeraj Rajesh&lt;/a&gt;, on &lt;a href=&quot;/publications/rajesh2026explainability/&quot;&gt;extending I/O observability to explainability&lt;/a&gt; with Sarah Neuwirth, &lt;a href=&quot;https://www.lofstead.org/&quot;&gt;Jay Lofstead&lt;/a&gt;, &lt;a href=&quot;https://ndryden.com/&quot;&gt;Nikoli Dryden&lt;/a&gt;, and &lt;a href=&quot;https://grc.iit.edu/center/team/hariharan-devarajan/&quot;&gt;Hariharan Devarajan&lt;/a&gt;, and &lt;a href=&quot;https://grc.iit.edu/center/team/rajni-pawar/&quot;&gt;Rajni Pawar&lt;/a&gt;, on &lt;a href=&quot;/publications/pawar2026gnnpress/&quot;&gt;compressed, bit-exact feature streaming&lt;/a&gt; for GNN training that outgrows HBM.&lt;/p&gt; &lt;h2 id=&quot;agents-that-watch-the-science&quot;&gt;Agents that watch the science&lt;/h2&gt; &lt;p&gt;On Wednesday &lt;a href=&quot;https://grc.iit.edu/center/team/hua-xu/&quot;&gt;Hua Xu&lt;/a&gt; presents &lt;a href=&quot;/publications/xu2026vigil/&quot;&gt;Vigil&lt;/a&gt; in the main track, with &lt;a href=&quot;https://grc.iit.edu/center/team/jaime-cernuda/&quot;&gt;Jaime Cernuda&lt;/a&gt;, &lt;a href=&quot;https://ana-gainaru.com&quot;&gt;Ana Gainaru&lt;/a&gt;, and &lt;a href=&quot;https://www.ornl.gov/staff-profile/scott-klasky&quot;&gt;Scott Klasky&lt;/a&gt;. The paper names a failure the stack never sees: simulations that run to completion while computing the wrong answer. Vigil pairs cheap science-aware triggers with a multimodal model that renders and judges the flagged data, and it can stop or restart the run from a checkpoint. Across four codes it detects every deliberately configured failure with zero false positives. Hua’s PDSW work-in-progress talk on &lt;a href=&quot;/publications/xu2026triggergated/&quot;&gt;trigger-gated agents&lt;/a&gt; and J. Teves’s talk on &lt;a href=&quot;/publications/teves2026typesafe/&quot;&gt;type-safe, dynamically defined metadata&lt;/a&gt; round out Monday.&lt;/p&gt; &lt;h2 id=&quot;thursday-hdf5-and-trustworthy-data&quot;&gt;Thursday: HDF5 and trustworthy data&lt;/h2&gt; &lt;p&gt;On Thursday, November 19, from 12:15 to 1:15pm CST in room W185d, I co-lead the Birds of a Feather session &lt;a href=&quot;https://sc26.conference-program.com/presentation/?id=bof168&amp;amp;sess=sess428&quot;&gt;Trustworthy Data at Scale: HDF5 in the Age of AI-Driven Science&lt;/a&gt;. Michael Breitenfeld of &lt;a href=&quot;https://www.hdfgroup.org/&quot;&gt;The HDF Group&lt;/a&gt; leads it, with Elena Pourmal of LifeBoat, &lt;a href=&quot;https://sbyna.github.io&quot;&gt;Suren Byna&lt;/a&gt; of Ohio State, and Quincey Koziol of NVIDIA. Domain scientists give short talks, and an opt-in agentic survey of attendee priorities steers the discussion and feeds HDF5 roadmap planning. Two of Monday’s PDSW papers work on HDF5 directly, and I want to hear what the community needs from it next.&lt;/p&gt; &lt;p&gt;I also serve on the SC26 State of the Practice Committee this year. If you are in Chicago, come find the students at their talks and posters, and come to W185d on Thursday.&lt;/p&gt; </content> </entry> <entry> <title type="html">Two PhDs, one architecture</title> <link rel="alternate" type="text/html" href="https://akougkas.io/journal/2025/two-phds-one-architecture/"/> <id>https://akougkas.io/journal/2025/two-phds-one-architecture/</id> <published>2025-12-15T00:00:00-06:00</published> <updated>2026-09-06T00:00:00-05:00</updated> <summary type="html">Luke Logan and Jaime Cernuda defended in the same semester. Their dissertations are the two halves of the runtime and the agent layer we are building, and they got there by owning the problems, not by executing mine.</summary> <content type="html" xml:base="https://akougkas.io/journal/2025/two-phds-one-architecture/">&lt;p&gt;&lt;a href=&quot;https://grc.iit.edu/center/team/luke-logan/&quot;&gt;Luke Logan&lt;/a&gt; defended in September. &lt;a href=&quot;https://grc.iit.edu/center/team/jaime-cernuda/&quot;&gt;Jaime Cernuda&lt;/a&gt; defended today. I co-advised both with &lt;a href=&quot;https://www.iit.edu/directory/people/xian-he-sun&quot;&gt;Xian-He Sun&lt;/a&gt;, and I want to write down what their two dissertations are, side by side, because together they are the architecture I have been describing on this site for a year.&lt;/p&gt; &lt;aside class=&quot;entry-glance&quot; aria-label=&quot;At a glance&quot;&gt; &lt;p class=&quot;entry-glance__kicker&quot;&gt;At a glance&lt;/p&gt; &lt;dl&gt; &lt;div&gt;&lt;dt&gt;Anchor&lt;/dt&gt;&lt;dd&gt;Two dissertation defenses, September and December 15, 2025&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;Systems&lt;/dt&gt;&lt;dd&gt;&lt;a href=&quot;/work/clio/&quot;&gt;CLIO&lt;/a&gt;, &lt;a href=&quot;/work/iowarp/&quot;&gt;IOWarp&lt;/a&gt;, &lt;a href=&quot;/work/hermes/&quot;&gt;Hermes&lt;/a&gt;&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;The claim&lt;/dt&gt;&lt;dd&gt;Two students who owned their own problems built the two halves of one architecture from opposite directions.&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;Code&lt;/dt&gt;&lt;dd&gt;&lt;a href=&quot;https://github.com/iowarp/clio-core&quot;&gt;iowarp/clio-core&lt;/a&gt;, &lt;a href=&quot;https://github.com/iowarp/clio-agent&quot;&gt;iowarp/clio-agent&lt;/a&gt;&lt;/dd&gt;&lt;/div&gt; &lt;/dl&gt; &lt;/aside&gt; &lt;h2 id=&quot;what-happened&quot;&gt;What happened&lt;/h2&gt; &lt;p&gt;&lt;a href=&quot;https://grc.iit.edu/center/team/luke-logan/&quot;&gt;Luke&lt;/a&gt; came to &lt;a href=&quot;https://www.iit.edu/&quot;&gt;Illinois Tech&lt;/a&gt; as an undergraduate researcher in 2017 and never really left the storage stack. His doctoral work runs through three systems.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;pMEMCPY.&lt;/strong&gt; A &lt;a href=&quot;/publications/logan2021pmemcpy/&quot;&gt;small I/O library&lt;/a&gt; that treats persistent memory as storage.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;LabStor, SC 2022.&lt;/strong&gt; A &lt;a href=&quot;/publications/logan2022labstor/&quot;&gt;modular userspace platform&lt;/a&gt; for building custom I/O stacks.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;MegaMmap, SC 2024.&lt;/strong&gt; A &lt;a href=&quot;/publications/logan2024megammap/&quot;&gt;software distributed shared memory&lt;/a&gt; that spills transparently across DRAM, NVMe, and the parallel file system, so an out-of-core algorithm can be written as if the data fit.&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;Underneath all three is the same instinct: separate what the data means from where it is placed, and let a runtime decide the placement. He also spent years as the engineer who kept &lt;a href=&quot;https://github.com/HDFGroup/hermes&quot;&gt;Hermes&lt;/a&gt; running, and the coroutine runtime he built for it became Chimaera.&lt;/p&gt; &lt;p&gt;&lt;a href=&quot;https://grc.iit.edu/center/team/jaime-cernuda/&quot;&gt;Jaime&lt;/a&gt;’s line runs through pipelines.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;HFlow, Cluster 2021.&lt;/strong&gt; Turned &lt;a href=&quot;/publications/cernuda2021hflow/&quot;&gt;data forwarding&lt;/a&gt; into ephemeral, relocatable tasks.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Hades, CCGrid 2024.&lt;/strong&gt; &lt;a href=&quot;/publications/cernuda2024hades/&quot;&gt;Derived quantities&lt;/a&gt; asynchronously during data production, so that analysis queries hit cached results instead of re-reading files.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;HStream, ICPP 2024.&lt;/strong&gt; &lt;a href=&quot;/publications/cernuda2024hstream/&quot;&gt;Separated the data plane&lt;/a&gt; from the compute plane at the task level, so a streaming engine could run out of core without thrashing.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Jarvis, PDSW 2024.&lt;/strong&gt; Made all of that &lt;a href=&quot;/publications/cernuda2024jarvis/&quot;&gt;reproducible to deploy&lt;/a&gt;.&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;His instinct is that a workflow is a graph of tasks with a data plane you have to manage on purpose.&lt;/p&gt; &lt;p&gt;Put those two instincts together and you have &lt;a href=&quot;/work/clio/&quot;&gt;CLIO&lt;/a&gt;. Luke’s placement runtime is &lt;a href=&quot;https://core.iowarp.ai/&quot;&gt;CLIO Core&lt;/a&gt;: the tiered context store whose public tags started landing in November. Jaime’s task-and-pipeline thinking is the shape of the agent and remote-execution side, where a science agent plans over tools and a relay submits work to a cluster and brings back logs, artifacts, and provenance. Neither of them set out to build a context layer for scientific AI. They built the two halves of one, from two different directions, and then we noticed. I would like to claim I planned it. I did not.&lt;/p&gt; &lt;aside class=&quot;entry-pull&quot;&gt; &lt;p&gt;The problems were theirs to define, and both of them ended up disagreeing with parts of my architecture for reasons I had not seen.&lt;/p&gt; &lt;/aside&gt; &lt;h2 id=&quot;what-i-am-taking-from-it&quot;&gt;What I am taking from it&lt;/h2&gt; &lt;p&gt;I did not assign either dissertation. I framed a research direction, &lt;a href=&quot;https://grc.iit.edu/research/projects/iowarp/&quot;&gt;IOWarp&lt;/a&gt; and what came before it, and then tried to make sure each of them owned a consequential problem inside it, with real interfaces, real collaborators, and real consequences when something broke. Luke owned the runtime. Jaime owned the pipeline.&lt;/p&gt; &lt;p&gt;The problems were theirs to define, and both of them ended up disagreeing with parts of my architecture for reasons I had not seen. That is the outcome I want from advising, and it is not the same thing as delegation.&lt;/p&gt; &lt;p&gt;I spent nearly nine years as an army officer, where delegation meant handing down a task with the answer attached. Advising is the opposite, and unlearning the habit took me longer than I would like to admit.&lt;/p&gt; &lt;p&gt;Two things I would tell a newer advisor.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;Keep a student on one line of work.&lt;/strong&gt; The students who stay on one line of work for five years produce systems, not only papers, and systems are what a research program can be built on after the paper deadline passes.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Make ownership visible outside the lab.&lt;/strong&gt; Luke’s name is on the runtime. Jaime’s name is on the streaming engine and the deployment tool. When someone at &lt;a href=&quot;https://www.anl.gov/&quot;&gt;Argonne&lt;/a&gt; or Livermore asks who built a thing, the answer is a person, not a group. I try to answer that way even when it costs me a line on my own CV.&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;The unresolved part is what happens next. A center that produced two systems builders in one semester has to decide whether it can keep them, and what it owes them if it does. I have opinions about that, but they belong in a later note once the decisions are real. Today I mostly feel older, in the good way.&lt;/p&gt; &lt;h2 id=&quot;update-september-2026&quot;&gt;Update, September 2026&lt;/h2&gt; &lt;p&gt;In January 2026 Luke and Jaime both joined the center as Assistant Research Professors. Luke leads day-to-day engineering of &lt;a href=&quot;/work/clio/&quot;&gt;CLIO Core&lt;/a&gt;, which reached &lt;a href=&quot;https://github.com/iowarp/clio-core/releases/tag/v2.2.1&quot;&gt;v2.2.1&lt;/a&gt; in August. Jaime leads day-to-day development of &lt;a href=&quot;https://github.com/iowarp/clio-agent&quot;&gt;CLIO Agent&lt;/a&gt;, at v0.9.1.1, and &lt;a href=&quot;https://github.com/iowarp/clio-relay&quot;&gt;CLIO Relay&lt;/a&gt;, at v1.6.8, the two components I described above as “the shape of the agent side.” The &lt;a href=&quot;/work/building-grc/&quot;&gt;Building GRC&lt;/a&gt; story carries their current roles.&lt;/p&gt; </content> </entry> <entry> <title type="html">Reflecting on SC25</title> <link rel="alternate" type="text/html" href="https://akougkas.io/journal/2025/reflecting-on-sc25/"/> <id>https://akougkas.io/journal/2025/reflecting-on-sc25/</id> <published>2025-11-22T00:00:00-06:00</published> <updated>2025-11-22T00:00:00-06:00</updated> <summary type="html">What I noticed at SC25 in St. Louis, what the group brought, what a year of running the PDSW program committee taught me about how a workshop gets made, and a few notes for people attending their first SC.</summary> <content type="html" xml:base="https://akougkas.io/journal/2025/reflecting-on-sc25/">&lt;p&gt;SC25 ran November 16 to 21 at America’s Center in St. Louis. I came home from Atlanta last year with a list of things to watch for. This year I came home with a different kind of list, because for most of 2025 I was on the other side of the table. The &lt;a href=&quot;https://www.pdsw.org/pdsw25/&quot;&gt;Parallel Data Systems Workshop&lt;/a&gt;, which I attended as a student and later presented at, was the workshop I was responsible for programming.&lt;/p&gt; &lt;aside class=&quot;entry-glance&quot; aria-label=&quot;At a glance&quot;&gt; &lt;p class=&quot;entry-glance__kicker&quot;&gt;At a glance&lt;/p&gt; &lt;dl&gt; &lt;div&gt;&lt;dt&gt;Anchor&lt;/dt&gt;&lt;dd&gt;SC25 in St. Louis, November 16 to 21, 2025, with PDSW on November 17&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;Systems&lt;/dt&gt;&lt;dd&gt;&lt;a href=&quot;/work/iowarp/&quot;&gt;IOWarp&lt;/a&gt;, &lt;a href=&quot;/work/deepio/&quot;&gt;DeepIO&lt;/a&gt;, &lt;a href=&quot;/work/building-grc/&quot;&gt;Building GRC&lt;/a&gt;&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;The claim&lt;/dt&gt;&lt;dd&gt;Programming a workshop resembles architecting a system: fix the interfaces early, then make sure the pieces meet.&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;Talk&lt;/dt&gt;&lt;dd&gt;&lt;a href=&quot;https://www.pdsw.org/pdsw25/&quot;&gt;PDSW 2025 program and committee&lt;/a&gt;&lt;/dd&gt;&lt;/div&gt; &lt;/dl&gt; &lt;/aside&gt; &lt;p&gt;I have been part of this community since 2013, the year I arrived at &lt;a href=&quot;https://www.iit.edu/&quot;&gt;Illinois Tech&lt;/a&gt;, and I decided early that the workshops were where the grown-ups were. This year I was, apparently, one of them.&lt;/p&gt; &lt;p&gt;St. Louis itself I can barely report on. The conference was the city, as it usually is, and what I saw of it was the inside of America’s Center and the corridor to the hotel.&lt;/p&gt; &lt;h2 id=&quot;the-workshop-side-of-the-table&quot;&gt;The workshop side of the table&lt;/h2&gt; &lt;p&gt;PDSW 2025 met on Monday, November 17. &lt;a href=&quot;https://sbyna.github.io&quot;&gt;Suren Byna&lt;/a&gt; was General Chair. Sarah Neuwirth and I were Program Co-Chairs. Ricardo Macedo ran reproducibility, and &lt;a href=&quot;https://grc.iit.edu/center/team/izzet-yildirim/&quot;&gt;Izzet Yildirim&lt;/a&gt;, a PhD student in our group, ran publicity. Rob Ross from &lt;a href=&quot;https://www.anl.gov/&quot;&gt;Argonne&lt;/a&gt; gave the invited talk, on how data management software supports science, which is the sentence I would use to describe what our whole field is for.&lt;/p&gt; &lt;p&gt;Here is what a year of that work looks like from the inside. A workshop proposal to the SC committee in February. A committee to form by April. A call for papers in May.&lt;/p&gt; &lt;p&gt;Full papers due August 8, with artifact descriptions a week later, because the workshop aligned with the SC25 Reproducibility Initiative and asked authors to structure artifacts so a reader can follow them years from now. Reviews through August, then a virtual program committee meeting in early September. Notifications September 5 and camera-ready by September 27. Work-in-progress abstracts through September 19, since a workshop is also supposed to be a place for ideas that are not papers yet.&lt;/p&gt; &lt;figure class=&quot;figure--diagram&quot;&gt; &lt;svg viewBox=&quot;0 0 640 240&quot; role=&quot;img&quot; aria-labelledby=&quot;fig-pdsw-year&quot;&gt; &lt;title id=&quot;fig-pdsw-year&quot;&gt;The program committee year for PDSW 2025, from the workshop proposal in February to the workshop day on November 17.&lt;/title&gt; &lt;text class=&quot;d-kicker&quot; x=&quot;24&quot; y=&quot;44&quot;&gt;february to november 2025&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;24&quot; y=&quot;96&quot;&gt;proposal&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;147&quot; y=&quot;132&quot; text-anchor=&quot;middle&quot;&gt;committee&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;209&quot; y=&quot;96&quot; text-anchor=&quot;middle&quot;&gt;call&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;409&quot; y=&quot;132&quot; text-anchor=&quot;middle&quot;&gt;papers due&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;465&quot; y=&quot;96&quot; text-anchor=&quot;middle&quot;&gt;notify&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;530&quot; y=&quot;132&quot; text-anchor=&quot;middle&quot;&gt;camera-ready&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;616&quot; y=&quot;96&quot; text-anchor=&quot;end&quot;&gt;workshop&lt;/text&gt; &lt;path class=&quot;d-rule&quot; d=&quot;M24 104 V168 M147 140 V168 M209 104 V168 M409 140 V168 M465 104 V168 M511 140 V168 M613 104 V168&quot; /&gt; &lt;path class=&quot;d-rule&quot; d=&quot;M24 176 H616&quot; /&gt; &lt;path class=&quot;d-line&quot; d=&quot;M24 168 V184 M147 168 V184 M209 168 V184 M409 168 V184 M465 168 V184 M511 168 V184 M613 168 V184&quot; /&gt; &lt;circle class=&quot;d-accent-fill&quot; cx=&quot;613&quot; cy=&quot;176&quot; r=&quot;6&quot; /&gt; &lt;/svg&gt; &lt;figcaption&gt;A single timeline with seven marks: the proposal in February, the committee by April, the call in May, papers due August 8, notifications September 5, camera-ready September 27, and the workshop on November 17.&lt;/figcaption&gt; &lt;/figure&gt; &lt;p&gt;Three things I took from running the committee.&lt;/p&gt; &lt;p&gt;&lt;strong&gt;Sort by consensus, then by variance.&lt;/strong&gt; The papers everyone agreed on, high or low, needed a few minutes each. The papers where three reviewers disagreed were the meeting. We ordered the discussion that way on purpose, and it turned a long call into a focused one.&lt;/p&gt; &lt;p&gt;&lt;strong&gt;Say the conflict rule out loud, every time.&lt;/strong&gt; Before each paper, the chair asks whether anyone has a conflict: same institution, advisor and advisee at any point, a recent collaboration, family, money. A conflicted person leaves the call for that discussion. That includes chairs. When a chair is conflicted, the co-chair runs the paper. Our group had a paper in the pool this year, so I am glad the protocol was written down in January rather than improvised in September.&lt;/p&gt; &lt;p&gt;I also sat on the technical program committee of the State of the Practice track this year and reviewed papers for it. That is the quieter side of the same job, and it taught me what a well-run track expects from its reviewers before the chairs ever meet.&lt;/p&gt; &lt;p&gt;&lt;strong&gt;A lead reviewer who can summarize in two minutes is worth more than a fourth review.&lt;/strong&gt; Contribution, strengths, weaknesses, recommendation. Then the others speak. My meeting guide said “two minutes” and I underestimated how much that one constraint would do. I have run enough meetings, in uniform and out of it, to know that the constraint is the meeting.&lt;/p&gt; &lt;p&gt;What I keep from the day itself is Rob’s talk and the hour after it. The title was a plain sentence about how data management software supports science, and the hallway afterwards turned it into three separate arguments about what that means once the consumer is an agent. I did not steer any of them. That is what a workshop is for.&lt;/p&gt; &lt;h2 id=&quot;what-the-group-brought&quot;&gt;What the group brought&lt;/h2&gt; &lt;p&gt;&lt;a href=&quot;https://grc.iit.edu/center/team/meng-tang/&quot;&gt;Meng Tang&lt;/a&gt; presented &lt;a href=&quot;/publications/tang2025awss3/&quot;&gt;Quantifying AWS S3 I/O Performance Boundaries Using the Roofline Model&lt;/a&gt; at PDSW, work with collaborators at &lt;a href=&quot;https://www.pnnl.gov/&quot;&gt;Pacific Northwest&lt;/a&gt; and Mainz. The paper extends roofline modeling to three ways of integrating S3 into an HPC workflow and finds up to a 6.4x difference in achievable bandwidth between them. A familiar model applied to an unfamiliar storage target, with numbers a practitioner can act on. That is the kind of paper I like to see from our group: a known tool pointed at something people actually pay for.&lt;/p&gt; &lt;p&gt;The conference marked one year since &lt;a href=&quot;https://grc.iit.edu/center/team/luke-logan/&quot;&gt;Luke Logan&lt;/a&gt;, who defended his dissertation in September, presented &lt;a href=&quot;/publications/logan2024megammap/&quot;&gt;MegaMmap&lt;/a&gt; at SC24. The idea, a memory abstraction that tiers transparently across DRAM, NVMe, and disk, has aged well. The pressure that motivated it in 2024 is stronger now, because inference keeps more state in memory than training ever did. The transfer engine inside &lt;a href=&quot;/work/iowarp/&quot;&gt;IOWarp&lt;/a&gt; is where that design continues.&lt;/p&gt; &lt;aside class=&quot;entry-pull&quot;&gt; &lt;p&gt;I have run enough meetings, in uniform and out of it, to know that the constraint is the meeting.&lt;/p&gt; &lt;/aside&gt; &lt;h2 id=&quot;what-i-noticed&quot;&gt;What I noticed&lt;/h2&gt; &lt;p&gt;On the floor, the change from last year was physical. Inference hardware was everywhere: racks built to serve models stood where storage vendors used to put their demo arrays, and the storage vendors who were there talked about feeding accelerators rather than about capacity. I am not going to name booths. The exhibit hall now looks like the argument I have been making in talks, and I found that more sobering than satisfying.&lt;/p&gt; &lt;p&gt;What I can say from the program alone is that the data and storage workshops have become the place where AI infrastructure questions get asked with numbers attached. This year’s PDSW papers covered layer-wise checkpointing for large language models, write isolation for in-memory databases on flexible data placement SSDs, scalable metadata for parallel I/O libraries, runtime I/O optimization, reinforcement learning dropped into system tuning, and secure in-storage execution of visualization workloads. The work-in-progress talks added lessons from 85,000 training checkpoints, &lt;a href=&quot;https://icl.utk.edu/papi/&quot;&gt;PAPI&lt;/a&gt; on AI accelerators, and &lt;a href=&quot;https://docs.daos.io/&quot;&gt;DAOS&lt;/a&gt; on 400 gigabit fabrics.&lt;/p&gt; &lt;p&gt;The panel, moderated by &lt;a href=&quot;https://ndryden.com/&quot;&gt;Nikoli Dryden&lt;/a&gt;, put a storage vendor and the community in one room to argue about where cloud and HPC storage should meet. None of these are new topics. All of them have new constraints because of what sits on the other end of the storage system now. That is the shift I described from Atlanta last year, and this time the community was measuring it rather than announcing it. I prefer the measuring. Announcing is what I did last year.&lt;/p&gt; &lt;h2 id=&quot;if-it-is-your-first-sc&quot;&gt;If it is your first SC&lt;/h2&gt; &lt;p&gt;Last year I wrote about pacing and networking. Those still hold. A few things I would add after a year in a committee seat.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;Go to a &lt;a href=&quot;https://sc25.supercomputing.org/program/workshops/&quot;&gt;workshop&lt;/a&gt; on Sunday or Monday.&lt;/strong&gt; The rooms are smaller, the questions are better, and the people who run the community are in them. The main technical program is where you watch. The workshops are where you talk.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Submit a work-in-progress abstract.&lt;/strong&gt; PDSW and most SC workshops have a WIP track with a later deadline and a lower bar. Five minutes in front of a room of people in your subfield does more for a first-year student than a poster in the main hall.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Volunteer for a committee role.&lt;/strong&gt; Publicity, web, artifact evaluation. &lt;a href=&quot;https://grc.iit.edu/center/team/izzet-yildirim/&quot;&gt;Izzet&lt;/a&gt; did it this year. You learn how decisions get made, and the people making them learn your name. I waited too long to do this myself, and I tell my students not to.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Read the artifact descriptions.&lt;/strong&gt; Since SC24 the reproducibility appendices have gotten good. They tell you more about how a system actually runs than the paper does.&lt;/li&gt; &lt;/ul&gt; &lt;h2 id=&quot;what-i-am-carrying-forward&quot;&gt;What I am carrying forward&lt;/h2&gt; &lt;p&gt;The thing I did not expect was how much programming a workshop resembles architecting a system. You fix the interfaces early, the call, the dates, the review form, and then most of the work is making sure the pieces meet. I should have known that. I build the systems version of it for a living. I want that discipline in how we run reviews inside the group, and I want more of our students on committees next year.&lt;/p&gt; &lt;p&gt;What I want out of SC26 is specific. All eight of my PhD students in the building, each with something to present: papers, posters, work-in-progress talks, a BoF. And demos ready for the whole &lt;a href=&quot;/work/clio/&quot;&gt;CLIO&lt;/a&gt; ecosystem, because I intend to treat SC26 as its soft launch.&lt;/p&gt; </content> </entry> <entry> <title type="html">Vibe Coding Fridays</title> <link rel="alternate" type="text/html" href="https://akougkas.io/journal/2025/vibe-coding-fridays/"/> <id>https://akougkas.io/journal/2025/vibe-coding-fridays/</id> <published>2025-08-29T00:00:00-05:00</published> <updated>2026-09-06T00:00:00-05:00</updated> <summary type="html">Twelve Fridays, three hours each, teaching a research group to work with coding agents: the frame that stuck, the bake-off at week five, the checklist, and what I would change.</summary> <content type="html" xml:base="https://akougkas.io/journal/2025/vibe-coding-fridays/">&lt;p&gt;The last session was today. Since June 6 the graduate researchers at the &lt;a href=&quot;/work/building-grc/&quot;&gt;Gnosis Research Center&lt;/a&gt; and I have spent every Friday from ten to one working with AI coding tools on real code, and I want to write down what the program was before the details fade.&lt;/p&gt; &lt;aside class=&quot;entry-glance&quot; aria-label=&quot;At a glance&quot;&gt; &lt;p class=&quot;entry-glance__kicker&quot;&gt;At a glance&lt;/p&gt; &lt;dl&gt; &lt;div&gt;&lt;dt&gt;Anchor&lt;/dt&gt;&lt;dd&gt;Twelve Friday sessions at the Gnosis Research Center, June 6 to August 29, 2025&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;Systems&lt;/dt&gt;&lt;dd&gt;&lt;a href=&quot;/work/building-grc/&quot;&gt;Building GRC&lt;/a&gt;, &lt;a href=&quot;/work/clio-coder/&quot;&gt;CLIO Coder&lt;/a&gt;&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;The claim&lt;/dt&gt;&lt;dd&gt;Using a coding agent well is a skill with structure, and a research group should learn it together, in the open.&lt;/dd&gt;&lt;/div&gt; &lt;div&gt;&lt;dt&gt;Code&lt;/dt&gt;&lt;dd&gt;&lt;a href=&quot;https://github.com/akougkas/claude-code-4-science&quot;&gt;claude-code-4-science&lt;/a&gt;, &lt;a href=&quot;https://github.com/akougkas/gemini-4-science&quot;&gt;gemini-4-science&lt;/a&gt;&lt;/dd&gt;&lt;/div&gt; &lt;/dl&gt; &lt;/aside&gt; &lt;p&gt;The premise was simple. The tools were changing faster than any course could, the students were already using them unevenly and mostly in secret, and I would rather teach the practice in the open than pretend it was not happening. So: twelve weeks, three hours a session, a tutorial, then hands-on work, then a recorded reflection. We called it Vibe Coding Fridays, half as a joke about the phrase and half because the name got people in the room. I still wince at the name a little. It worked.&lt;/p&gt; &lt;p&gt;June was foundations. Environment setup on the first day, with &lt;a href=&quot;https://www.docker.com/&quot;&gt;Docker&lt;/a&gt;, &lt;a href=&quot;https://code.visualstudio.com/&quot;&gt;VS Code&lt;/a&gt;, and &lt;a href=&quot;https://docs.astral.sh/uv/&quot;&gt;&lt;code class=&quot;language-plaintext highlighter-rouge&quot;&gt;uv&lt;/code&gt;&lt;/a&gt; as the only sanctioned tooling, and one rule for the summer: ask the AI first, then ask the group. That rule was mine. People pushed back in June and had stopped noticing by July.&lt;/p&gt; &lt;p&gt;Week two introduced what I called the three pillars of working with these tools. Everything after that was those three ideas applied to harder situations.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;Context is king.&lt;/strong&gt; The model is only as good as what you put in front of it.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Iterative refinement.&lt;/strong&gt; Start broad and narrow with specific feedback, never expect a one-shot.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Human-AI partnership.&lt;/strong&gt; The model generates, the person architects and decides.&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;July was individual mastery: command-line agents, local models, hooks, parallel sub-agents in &lt;a href=&quot;https://git-scm.com/docs/git-worktree&quot;&gt;Git worktrees&lt;/a&gt;, and then a week on building an agent that runs for thirty minutes without you. August was competitive: teams, rapid prototyping, and a final round of presentations. The two public artifacts that came out of that last stretch are &lt;a href=&quot;https://github.com/akougkas/claude-code-4-science&quot;&gt;Claude Code 4 Science&lt;/a&gt; and &lt;a href=&quot;https://github.com/akougkas/gemini-4-science&quot;&gt;Gemini 4 Science&lt;/a&gt;, scientific-computing extensions for the two terminal agents, both published in the third week of August.&lt;/p&gt; &lt;figure class=&quot;figure--diagram&quot;&gt; &lt;svg viewBox=&quot;0 0 640 240&quot; role=&quot;img&quot; aria-labelledby=&quot;fig-vcf-arc&quot;&gt; &lt;title id=&quot;fig-vcf-arc&quot;&gt;Twelve Fridays in three phases: foundations in June, individual mastery in July, competitive work in August, with the bake-off in week five.&lt;/title&gt; &lt;text class=&quot;d-kicker&quot; x=&quot;24&quot; y=&quot;44&quot;&gt;june&lt;/text&gt; &lt;text class=&quot;d-kicker&quot; x=&quot;225&quot; y=&quot;44&quot;&gt;july&lt;/text&gt; &lt;text class=&quot;d-kicker&quot; x=&quot;426&quot; y=&quot;44&quot;&gt;august&lt;/text&gt; &lt;rect class=&quot;d-soft&quot; x=&quot;24&quot; y=&quot;60&quot; width=&quot;190&quot; height=&quot;64&quot; rx=&quot;4&quot; /&gt; &lt;rect class=&quot;d-soft&quot; x=&quot;225&quot; y=&quot;60&quot; width=&quot;190&quot; height=&quot;64&quot; rx=&quot;4&quot; /&gt; &lt;rect class=&quot;d-box&quot; x=&quot;426&quot; y=&quot;60&quot; width=&quot;190&quot; height=&quot;64&quot; rx=&quot;4&quot; /&gt; &lt;text class=&quot;d-label&quot; x=&quot;119&quot; y=&quot;98&quot; text-anchor=&quot;middle&quot;&gt;foundations&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;320&quot; y=&quot;98&quot; text-anchor=&quot;middle&quot;&gt;individual mastery&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;521&quot; y=&quot;98&quot; text-anchor=&quot;middle&quot;&gt;competitive&lt;/text&gt; &lt;path class=&quot;d-rule&quot; d=&quot;M24 160 H616&quot; /&gt; &lt;path class=&quot;d-line&quot; d=&quot;M24 152 V168 M616 152 V168&quot; /&gt; &lt;circle class=&quot;d-accent-fill&quot; cx=&quot;246&quot; cy=&quot;160&quot; r=&quot;6&quot; /&gt; &lt;text class=&quot;d-label&quot; x=&quot;24&quot; y=&quot;196&quot;&gt;week 1&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;246&quot; y=&quot;196&quot; text-anchor=&quot;middle&quot;&gt;bake-off, week 5&lt;/text&gt; &lt;text class=&quot;d-label&quot; x=&quot;616&quot; y=&quot;196&quot; text-anchor=&quot;end&quot;&gt;week 12&lt;/text&gt; &lt;/svg&gt; &lt;figcaption&gt;A twelve-week line in three blocks, June for foundations, July for individual mastery, August for competitive work, with a marker at week five for the bake-off.&lt;/figcaption&gt; &lt;/figure&gt; &lt;h2 id=&quot;prompts-are-programs&quot;&gt;Prompts are programs&lt;/h2&gt; &lt;p&gt;The frame from week three is the one I would keep if I could keep only one. A prompt is a function.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;The system context&lt;/strong&gt; is its global state.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;The inputs&lt;/strong&gt; are its arguments: file paths, error logs, an API signature.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;The body&lt;/strong&gt; is the logic, best written as numbered steps.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;The output format&lt;/strong&gt; is the return type.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;The fallbacks&lt;/strong&gt; are the error handling: what the model should do when an input is missing or a step fails.&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;Once people saw prompts that way, two things happened. They started keeping a prompt library under version control, with templates that take `` and carry their own test cases. And they stopped writing “fix the module.” The week’s exercise was a four-step chain, requirements to architecture to code to tests, where each prompt’s output was the next prompt’s input, and the discipline of writing down the interface between steps did more for prompt quality than any list of tips. I would teach this frame before I taught anything about models. It is the one idea from the summer I have not had to revise.&lt;/p&gt; &lt;aside class=&quot;entry-pull&quot;&gt; &lt;p&gt;A prompt is a function.&lt;/p&gt; &lt;/aside&gt; &lt;h2 id=&quot;the-bake-off&quot;&gt;The bake-off&lt;/h2&gt; &lt;p&gt;Week five was the part people still talk about. Everyone picked a command-line agent, &lt;a href=&quot;https://github.com/anthropics/claude-code&quot;&gt;Claude Code&lt;/a&gt;, &lt;a href=&quot;https://github.com/google-gemini/gemini-cli&quot;&gt;Gemini CLI&lt;/a&gt;, &lt;a href=&quot;https://github.com/openai/codex&quot;&gt;Codex CLI&lt;/a&gt;, or a local model served through &lt;a href=&quot;https://ollama.com/&quot;&gt;Ollama&lt;/a&gt;, and ran the same ninety-minute workflow on the same deliberately incomplete FastAPI service: analyze the repository, add a feature with tests and docs, review and optimize existing code, generate documentation, and set up a container and a pipeline. Scoring covered completion, code quality, creative use of the tool, and resource efficiency, with a separate prize for the best result from a local model on limited hardware, because I wanted the room to take the local path seriously.&lt;/p&gt; &lt;p&gt;I am not going to publish a ranking from that afternoon, because there is none worth publishing. We scored on judgment, in the room, and the argument afterwards was about the rubric more than about the code. What I can say is that the local-model entries made the case for the resource-efficiency prize better than my slide did: a model on one workstation finished real parts of the workflow while spending nothing outside the building, and nobody in the room had expected that. The lesson I kept is the one below, under what I would change: write the rubric first.&lt;/p&gt; &lt;h2 id=&quot;the-checklist&quot;&gt;The checklist&lt;/h2&gt; &lt;p&gt;Week six asked everyone to build an agent that runs unattended for half an hour with clear logs. The framing was three words: programmable, observable, reproducible. Decisions and loops; logs, traces, and usage counts; configuration files and Git worktrees so a run can be repeated. Before anyone called an agent done we went through a short list:&lt;/p&gt; &lt;ul&gt; &lt;li&gt;a kill switch&lt;/li&gt; &lt;li&gt;rate limiting&lt;/li&gt; &lt;li&gt;timestamped logs&lt;/li&gt; &lt;li&gt;configuration outside the code&lt;/li&gt; &lt;li&gt;error recovery&lt;/li&gt; &lt;li&gt;human approval gates where the action is consequential&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;That list is short on purpose. I wrote it the way the army taught me to write checklists: short enough that nobody skips it. It is also the list I would hold any agent to before letting it run unattended.&lt;/p&gt; &lt;h2 id=&quot;what-i-would-change&quot;&gt;What I would change&lt;/h2&gt; &lt;p&gt;Most of these are mistakes I would have caught in a course I had taught before.&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;Start in the terminal, not the editor.&lt;/strong&gt; The first weeks used IDE assistants and a web-app exercise, and the group only became fluent once the agents moved to the command line and could run things.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Write the rubric before the bake-off.&lt;/strong&gt; We scored on judgment, and the arguments afterwards were about the rubric, not the code.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Eight weeks, not twelve.&lt;/strong&gt; Twelve is long for a summer; eight would hold attention better with the same arc.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;Record the reflections properly.&lt;/strong&gt; The fifteen minutes at the end of each session, where people said what worked and what did not, were the most valuable part, and I have the least of it written down.&lt;/li&gt; &lt;/ul&gt; &lt;p&gt;The seminar continues into the fall in a smaller form. The open question is whether the practice survives the summer, or whether it needs the Friday room to keep going. My guess is that it needs the room, and that the room is cheap.&lt;/p&gt; &lt;h2 id=&quot;update-september-2026&quot;&gt;Update, September 2026&lt;/h2&gt; &lt;p&gt;That six-item checklist became, almost line for line, the safety net inside &lt;a href=&quot;/work/clio-coder/&quot;&gt;CLIO Coder&lt;/a&gt;, the coding agent I started the following April. The seminar did survive the summer, as the AI-assisted programming seminar &lt;a href=&quot;https://grc.iit.edu/&quot;&gt;the center&lt;/a&gt; still runs.&lt;/p&gt; </content> </entry> <entry> <title type="html">Reflecting on SC24</title> <link rel="alternate" type="text/html" href="https://akougkas.io/journal/2024/thoughts-on-sc24/"/> <id>https://akougkas.io/journal/2024/thoughts-on-sc24/</id> <published>2024-11-26T10:41:16-06:00</published> <updated>2026-08-26T00:00:00-05:00</updated> <summary type="html">What I noticed at SC24, what our group brought to Atlanta, and a few practical notes for first-time conference attendees.</summary> <content type="html" xml:base="https://akougkas.io/journal/2024/thoughts-on-sc24/">&lt;h2 id=&quot;my-experience-at-supercomputing-2024-sc24&quot;&gt;My Experience at Supercomputing 2024 (SC24)&lt;/h2&gt; &lt;p&gt;this year’s supercomputing conference (sc24) in atlanta was a whirlwind! with 17,000 attendees, it was a massive gathering of the high-performance computing (hpc) community, which i’ve been a part of since 2013. this year felt particularly special, not just because of the sheer scale of the event, but also because of the significant shift towards ai and the prominent role my team and i played.&lt;/p&gt; &lt;figure&gt; &lt;img src=&quot;/assets/img/sc24.webp&quot; alt=&quot;Members of the Gnosis Research Center together at SC24 in Atlanta&quot; width=&quot;3963&quot; height=&quot;2229&quot; loading=&quot;eager&quot; /&gt; &lt;figcaption&gt;gnosis research center at sc24 (missing some people).&lt;/figcaption&gt; &lt;/figure&gt; &lt;h2 id=&quot;the-gnosis-research-center-takes-on-sc24&quot;&gt;The Gnosis Research Center Takes on SC24&lt;/h2&gt; &lt;p&gt;sc24 was a major event for the gnosis research center, which i co-founded and currently direct. i was joined by five of my students and a fellow faculty member, and we were thrilled to have a strong presence at the conference. we presented papers, held posters, led workshops, participated in panels, and had a fantastic time networking with colleagues and peers.&lt;/p&gt; &lt;p&gt;two papers anchored part of that program: &lt;a href=&quot;/publications/logan2024megammap/&quot;&gt;megammap&lt;/a&gt;, which explores the boundary between memory and storage for data-intensive workloads, and &lt;a href=&quot;/publications/devarajan2024dftracer/&quot;&gt;dftracer&lt;/a&gt;, which makes workflow data movement easier to trace and analyze. the publication records carry the authorship, abstracts, artifacts, and durable citation details; what mattered to me in atlanta was watching those ideas become conversations with the community.&lt;/p&gt; &lt;p&gt;i personally had the honor of co-chairing the &lt;a href=&quot;https://www.pdsw.org/pdsw24/&quot;&gt;pdsw workshop&lt;/a&gt; on parallel data storage, a topic that’s incredibly important in the era of ai and big data. i also presented our &lt;a href=&quot;/work/iowarp/&quot;&gt;iowarp project&lt;/a&gt; at the &lt;a href=&quot;https://www.hdfgroup.org/hug/hug24/agenda/&quot;&gt;hdf5 user group meeting&lt;/a&gt;, showcasing some of the exciting research we’re doing at gnosis.&lt;/p&gt; &lt;p&gt;seeing my students actively participate and contribute to the conference was incredibly rewarding. we even had a poster nominated for the best poster award! it’s through events like sc24 that we can foster the next generation of researchers and innovators in hpc.&lt;/p&gt; &lt;h2 id=&quot;the-rise-of-ai-in-hpc&quot;&gt;The Rise of AI in HPC&lt;/h2&gt; &lt;p&gt;one of the most striking trends at sc24 was the undeniable convergence of ai and hpc. this is a natural evolution, as ai requires the immense computational power that hpc offers, and hpc can benefit greatly from the intelligence ai brings to scientific workflows.&lt;/p&gt; &lt;p&gt;this shift towards ai is a game-changer. for the first time, we have tools that can generate new content, act as intelligent assistants, and automate tedious tasks. this opens up incredible possibilities for scientific discovery and exploration.&lt;/p&gt; &lt;p&gt;imagine, for instance, using ai in climate science. we can feed ai massive datasets from sensors, satellites, and simulations. the ai can then process this information, generate insights, and even formulate hypotheses for the next iteration of climate modeling. this is just one example of how ai is becoming deeply ingrained in scientific workflows, both for learning and inference.&lt;/p&gt; &lt;p&gt;the growing presence of major players like nvidia, google, and microsoft at sc24 further underscores this trend. these companies are pushing the boundaries of ai and hpc, and their engagement with academic institutions and research labs is fueling innovation in the field.&lt;/p&gt; &lt;h2 id=&quot;supercomputing-goes-mainstream&quot;&gt;Supercomputing Goes Mainstream&lt;/h2&gt; &lt;p&gt;another key takeaway from sc24 was the sense that hpc is finally breaking out of its niche and entering the mainstream. with the rise of ai, the need for supercomputing resources is becoming more widespread, and this is attracting attention from a broader audience.&lt;/p&gt; &lt;p&gt;this is a positive development for several reasons. first, it exposes more people to the exciting world of hpc, potentially attracting new talent to the field. we desperately need more skilled researchers and engineers in hpc, and this increased visibility can help address the talent gap.&lt;/p&gt; &lt;p&gt;second, a wider audience means more scrutiny and diverse perspectives, which can lead to innovative solutions and accelerate progress. as hpc becomes more critical to various industries and applications, the collective effort to overcome challenges and push the boundaries of what’s possible will only intensify.&lt;/p&gt; &lt;h2 id=&quot;tips-and-hints-for-conference-goers&quot;&gt;Tips and Hints for Conference Goers&lt;/h2&gt; &lt;p&gt;attending a conference like sc24 can be overwhelming, especially for newcomers. here are a few tips to help you make the most of your experience:&lt;/p&gt; &lt;ul&gt; &lt;li&gt;&lt;strong&gt;prioritize networking:&lt;/strong&gt; focus on events where you can interact with people, such as workshops, panels, and social gatherings. don’t get bogged down in technical talks; those are usually available in the proceedings afterwards. coffee breaks are a great opportunity to strike up conversations and meet new people.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;pace yourself:&lt;/strong&gt; sc is a marathon, not a sprint. get plenty of rest, stay hydrated, and bring healthy snacks to keep your energy levels up.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;be curious:&lt;/strong&gt; don’t be afraid to approach people and ask about their work. most people are passionate about their research and happy to share their insights.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;come prepared:&lt;/strong&gt; bring business cards to exchange contact information quickly and easily.&lt;/li&gt; &lt;li&gt;&lt;strong&gt;promote your work:&lt;/strong&gt; if you’re presenting at the conference, use social media to generate interest and draw people to your session.&lt;/li&gt; &lt;/ul&gt; &lt;h2 id=&quot;looking-ahead&quot;&gt;Looking Ahead&lt;/h2&gt; &lt;p&gt;sc24 was a fantastic experience, and i left feeling incredibly energized about the future of hpc and ai. there are still challenges to overcome, particularly in the area of data management for ai workloads. however, the energy and innovation on display at sc24 give me confidence that we’re on the right track.&lt;/p&gt; &lt;p&gt;i’m already looking forward to sc25!&lt;/p&gt; </content> </entry> </feed>