Meteorologist, oceanographer, Python solvem probler, experimenter, Raspberry Pi tinkerer, geek dad, manager, space enthusiast, ham radio lover, science columnist, ADHDer, synthetic intellect curious, linux admin cosplayer.

COLA: The Center for Ocean-Land-Atmosphere Studies — Founding and Evolution

Research notes for a long-form blog post on COLA’s institutional history, the architectural pivot from supercomputer to workstation cluster, and the divergence between two operational philosophies in late-twentieth-century atmospheric science.

Pre-history: From Banaras to Greenbelt (1944-1979)

COLA cannot be understood without the biography of its founder. Jagadish Shukla was born in 1944 in Mirdha (sometimes given as “Middha”), Ballia district, Uttar Pradesh, India, in a village without electricity, plumbing, or paved roads. He attended primary school under a banyan tree – a fact he has cited frequently in subsequent autobiographical writing, including his 2024 memoir A Billion Butterflies: A Life in Climate and Chaos Theory.1 He earned a B.Sc. (honours) in physics, mathematics, and geology from Banaras Hindu University, an M.Sc. in geophysics in 1964, and a Ph.D. in geophysics from BHU in 1971. In 1970, at age 26, he arrived at MIT on an F-1 student visa and completed an Sc.D. in meteorology in 1976 under Jule Gregory Charney – the same Charney who was Mark Cane’s advisor a decade earlier (covered in Post 27) and the same Charney whose 1950 ENIAC forecast was the founding moment of computational meteorology (covered in Post 02).2

After his MIT doctorate Shukla spent time at Princeton’s GFDL before joining NASA’s Goddard Space Flight Center in Greenbelt, Maryland in 1979, where he became head of the climate modeling group. At Goddard he ran sensitivity experiments with the GFDL-derived GCM, integrating it with prescribed sea-surface-temperature anomalies over the Arabian Sea to test his developing argument that the seasonal mean state of the tropical atmosphere was not sensitively dependent on initial conditions in the way that day-to-day weather was. The 1981 paper that came out of this work – Shukla, J. (1981) “Dynamical Predictability of Monthly Means,” Journal of the Atmospheric Sciences 38(12), 2547-2572 – ran sixty-day integrations of a global GCM with nine different initial conditions but identical boundary conditions (SST, snow, sea ice, soil moisture). It found that “the evolution of long waves remains sufficiently predictable at least up to one month, and possibly up to 45 days, and the Northern Hemisphere appears to be more predictable than the Southern Hemisphere.”3

In parallel he co-authored, with his Ph.D. supervisor, the foundational theoretical statement of the field: Charney, J. G., and J. Shukla, 1981: “Predictability of monsoons,” in Monsoon Dynamics, J. Lighthill and R. P. Pearce, eds., Cambridge University Press, pp. 99-108. The Charney-Shukla paper argued that low-latitude variability was driven principally by slowly varying surface boundary forcing – ocean temperature, soil moisture, snow cover, vegetation – rather than by the fast atmospheric instabilities that govern mid-latitude weather. This is the Charney-Shukla hypothesis, which would be the central paradigm for monsoon predictability research for the subsequent thirty-five years.4

In 1982 Shukla received the NASA Exceptional Scientific Achievement Medal, the highest civilian honour the agency awards, for this work.

COLA at Maryland (1984-1992)

In 1984 Shukla left NASA Goddard to join the University of Maryland Department of Meteorology as Distinguished Professor. The same year he and several colleagues – principally James Kinter III, Edwin K. Schneider, David M. Straus, and Paul A. Dirmeyer – established the Center for Ocean-Land-Atmosphere Interactions within the University of Maryland Department of Meteorology. (The original name varied somewhat in early documents; “Studies” replaced “Interactions” by the late 1980s.) The institutional argument was that the existing structure of American climate-research funding – programme grants for individual investigators, with computing time bought from NCAR or from one of the Department of Energy laboratories – was not adequate for the systematic exploration of seasonal predictability that the Charney-Shukla hypothesis required. Sensitivity experiments needed to be run in ensembles of many tens or hundreds; each ensemble member required a multi-month GCM integration; and the small number of national supercomputing centres could not allocate the throughput required.5

Maryland’s computing infrastructure in the late 1980s consisted, like every American research university’s of the period, of a mixture of departmental Sun and DEC workstations, a campus IBM mainframe for administrative computing, and contracted access to NCAR’s Cray for the senior meteorology faculty. The COLA group used this conventional pattern for most of its first decade. The Maryland years were, in the language of the post-1993 institutional history, the period of demonstrating that the science could be done without a Cray of one’s own.

The Maryland-era COLA produced a substantial body of work. James Kinter – born and educated in the United States, with a 1984 Ph.D. in geophysical fluid dynamics from Princeton University and a National Research Council postdoctoral position at NASA Goddard before joining COLA – led much of the seasonal-prediction model development. Schneider and Straus contributed atmospheric dynamics; Dirmeyer focused on land-surface processes. By the early 1990s the COLA group had become recognisably distinct from both the Maryland Department of Meteorology that hosted it and from the NASA Goddard climate-modeling group from which Shukla had come.6

The 1992-1993 institutional pivot — IGES and the new COLA

In late 1992 and early 1993 Shukla took what his Wikipedia biographer would later describe as “an extraordinary and risky decision”: he resigned his tenured professorship at the University of Maryland and incorporated, as a private Maryland nonprofit corporation, the Institute of Global Environment and Society (IGES). IGES was founded in January 1993; the corporation’s first NSF grants – totalling $1.7 million – were dated 6 January 1993.7 Shukla named himself President of IGES; the entire former Maryland COLA group – Kinter, Schneider, Straus, Dirmeyer – transitioned to IGES in 1993 as the staff of its principal research unit, now formally named the Center for Ocean-Land-Atmosphere Studies.

The institution was located in Calverton, Maryland – a Washington suburb in Prince George’s County, about ten miles from the University of Maryland’s College Park campus, in a small commercial office park. (Apocryphally, Shukla started IGES “in his garage”; the institutional record shows that the first formal IGES office was in Calverton, but the garage characterisation has appeared in interviews and in the Wikipedia biography.)

In 1994 – the second year of IGES’s existence – COLA’s founding director secured the institutional bedrock that would define the organisation for the next twenty years: a five-year block grant of approximately $2.25 million per year for 1994-1998, jointly funded by NSF, NOAA, and NASA. This was the first of four consecutive five-year block grants; subsequent renewals followed in 1999, 2004, and 2009, each at higher levels. The total federal investment in IGES/COLA over its independent existence (1993-2014) is estimated by independent sources at over $75 million in research grants.8

Shukla also took a faculty appointment at George Mason University in 1994, which set up the eventual institutional move two decades later. But for the period 1993-2013, COLA’s primary institutional home was IGES, and its principal address was Calverton, Maryland.

The institutional argument for IGES, as it appears in the surviving 1993 documents and in Shukla’s later writing, was three-pointed: (1) seasonal-prediction research required scale of computing throughput that the standard university research-grant model could not deliver; (2) it required a stable multi-year staff that the standard postdoctoral-rotation model of academic departments could not provide; (3) it required computing infrastructure that could be owned, configured, and operated specifically for the kind of ensemble-experiment workload that the Charney-Shukla hypothesis required to be tested.

The architectural pivot — from supercomputer to workstation cluster

The third of these arguments is the technically load-bearing one. In 1993 the dominant pattern in American atmospheric-science computing was Cray big iron: the Cray X-MP and Y-MP at NCAR, the Cray C90 at Oak Ridge, the Cray X-MP/48 at ECMWF (covered in Post 34). A new Cray Y-MP/8 cost on the order of $20-30 million; a Cray C90 cost about $30-40 million. The institutional model was: a national centre owned the machine, and individual investigators bought time on it through programme allocations that ran in the hundreds-of-CPU-hours range per year per principal investigator.9

This model was wrong for the kind of work IGES/COLA needed to do. Systematic seasonal-predictability experiments required thousands of model integrations of months of simulated time each. A 100-hour annual NCAR allocation – the upper end of what an individual investigator could expect – bought you maybe ten or twenty integrations on a Cray Y-MP. That was not enough to test the Charney-Shukla hypothesis with any statistical power.

Shukla’s institutional bet, made concrete in 1993-1995, was that commodity workstation clusters – the architectural pattern that Donald Becker and Thomas Sterling had invented at NASA Goddard’s CESDIS in summer 1994 (about which more below) – would let a small institution own its own throughput at one to two orders of magnitude lower capital cost than a Cray. The economics were stark: a 16-node cluster of commodity Pentium-class workstations cost on the order of $100,000 to $200,000 in 1994-1996 dollars. A 64-node cluster cost about $500,000 to $1 million. Either was within reach of a $2.25-million-per-year block grant. Both were one to two orders of magnitude cheaper than a Cray, and – for the kind of workload that ensemble-experiment seasonal-prediction research generated – delivered comparable aggregate throughput.10

The COLA hardware path through the 1990s and 2000s reflected this architectural commitment. Through the mid-1990s the principal compute platform was a mixture of SGI workstations and IBM RS/6000 systems running proprietary Unix – the conventional research-station hardware of the period. From 1996-1997 onwards, following the rapid adoption of the Beowulf pattern across NASA Goddard and the rest of the atmospheric-science community, COLA transitioned to commodity Pentium-class Linux clusters built on the Becker-Sterling Beowulf model. By 2000 the bulk of COLA’s seasonal-prediction throughput was running on x86 Linux clusters; by 2005 the SGI and IBM Unix systems had been retired in favour of larger Linux clusters; by 2010 the operational architecture was indistinguishable from that of any other major Beowulf-pattern installation in scientific computing.

The Beowulf pattern — Becker, Sterling, and the 1994 NASA Goddard prototype

The Beowulf pattern is the architectural background against which COLA’s institutional bet has to be read. Donald Becker and Thomas Sterling, working at the Center of Excellence in Space Data and Information Sciences (CESDIS) at NASA Goddard Space Flight Center in Greenbelt, Maryland – a few miles from where COLA was being founded the same year – built the first Beowulf cluster in summer 1994. The original prototype consisted of 16 Intel 80486 DX4 processors clocked at 100 MHz, each on its own commodity desktop motherboard, each with 16 MB of DRAM and a 500 MB hard drive, all connected by two channel-bonded 10-Mbps Ethernet networks. The aggregate system had 256 MB of memory and 8 GB of disk and ran approximately 500 megaflops sustained on scientific code – which was a substantial fraction of a contemporary mid-range supercomputer’s performance for a small fraction of the cost.11

The architectural philosophy of Beowulf was set out in the founding paper: Sterling, T. L., D. Savarese, D. J. Becker, J. E. Dorband, U. A. Ranawake, and C. V. Packer (1995) “BEOWULF: A Parallel Workstation for Scientific Computation,” Proceedings of the International Conference on Parallel Processing 1995, Vol. I: Architecture, pp. 11-14. The paper’s design statement reads: “The Beowulf research project has been initiated to explore the opportunity of exploiting Network-of-Workstation concepts to provide high performance single user workstation performance at exceptional cost.” The design “avoids the use of any custom components, choosing instead to leverage the performance to cost benefits not only of mass market chips but of manufactured subsystems as well.” The paper reported that on a CFD code (Prometheus), the 16-node Beowulf delivered 60 megaflops sustained versus 4.5 megaflops on a single processor, a 13.3x speedup with 19% scaling degradation; and that the Cray T3D (one of Cray’s first MIMD distributed-memory machines) “performed less than 2.5 times better than Beowulf for the same number of processors.”12

Sterling’s own retrospective on the resistance the project faced is vivid: “Not only did nobody care, but there were even a number of people hostile to this project.” And: “We broke all the rules, and there was tremendous resentment, and of course the vendors hated it.” NASA Goddard’s eventual recognition of Beowulf – the project was inducted into the Space Technology Hall of Fame in 2022, and NASA’s Spinoff magazine documented its broader scientific impact – came many years after the founding skepticism had been overcome. The design choices that mattered – Linux as the operating system (Becker had been writing custom Linux Ethernet drivers since 1992 and was, by 1994, one of the most prolific Linux kernel contributors), Ethernet as the interconnect, MPI/PVM as the message-passing library, and commodity x86 as the processor base – were the ones that became the operational standard of every subsequent climate-research cluster.13

The name “Beowulf” was, Sterling has admitted, almost arbitrary: when a Goddard administrator asked him for a project name on the spot, he reached for the Old English epic on his bookshelf because his mother had majored in Old English. (“Oh hell, just call it Beowulf.”)14

The Beowulf paradigm spread through atmospheric science with remarkable speed. By 1996-1997 there were Beowulf-class clusters at virtually every major US atmospheric-science centre: NASA Goddard had multiple installations; NCAR was running its first commodity-cluster experiments by 1997; Argonne and Oak Ridge had Linux-cluster climate-modeling pilots running by 1998; the major university atmospheric-science programmes – Maryland, GMU/COLA, Colorado State, Florida State, Texas A&M – had all built Beowulf-pattern clusters by 2000. NASA’s Spinoff magazine in 2020 explicitly noted that “Beowulf clusters were 10 times cheaper” than Cray supercomputers for the kind of workload they served, which is why “the global move away from expensive, proprietary supercomputer systems and toward the adoption of PC clusters and open-source software by the scientific community” took the form it did.15

COLA’s scientific output (1993-2010)

The 1990s and 2000s were COLA’s scientific peak as an independent institution. The 1998 paper that defined the field is Shukla, J. (1998) “Predictability in the Midst of Chaos: A Scientific Basis for Climate Forecasting,” Science 282(5389), 728-731 (published 23 October 1998).16 The paper, written from “Center for Ocean-Land-Atmosphere Studies, Institute of Global Environment and Society, Calverton, MD” (the affiliation address that defined COLA’s identity for the entire IGES period), made the case that “while the Earth’s atmosphere is generally considered to be chaotic and sensitively dependent on initial conditions, certain regions of the atmosphere are an exception, where wind patterns and rainfall in certain tropical regions are so strongly determined by sea surface temperature that they do not show sensitive dependence on initial conditions.” The implication was operational: “it should be possible to predict large-scale tropical circulation and rainfall for as long as ocean temperature can be predicted.”

The COLA atmospheric model – a triangular-truncation T42 spectral primitive-equation atmospheric GCM, dynamic core derived from the NCAR Community Climate Model version 3, with COLA-specific parameterisations for radiation, convection, and land-surface processes – became the workhorse for the institution’s seasonal-prediction work. The seasonal-prediction methodology was set out in Shukla, J., J. L. Anderson, D. M. Baumhefner, et al. (2000) “Dynamical Seasonal Prediction,” Bulletin of the American Meteorological Society 81(11), 2593-2606, which described 16 winter seasons of hindcast experiments (December-March 1981/82 through 1996/97) using the COLA model.17

In parallel, Ben Kirtman, Paul Schopf, and Edwin Schneider developed COLA’s coupled atmosphere-ocean modelling capability. The COLA anomaly coupled model, described in Kirtman, B. P., Y. Fan, and E. K. Schneider (2002) “The COLA Global Coupled and Anomaly Coupled Ocean-Atmosphere General Circulation Model,” Journal of Climate 15(17), 2301-2320 and in Kirtman, B. P. (2003) “The COLA Anomaly Coupled Model: Ensemble ENSO Prediction,” Monthly Weather Review 131(10), 2324-2341, combined the COLA atmospheric GCM with the GFDL Modular Ocean Model version 3.0. This was the operational ENSO-prediction model that COLA contributed to the IRI multi-model forecast plume from the late 1990s onwards.18

The international-collaboration component of COLA’s mandate was substantial. The institution helped establish climate-research programmes in India (where Shukla was instrumental in modernising India’s weather-forecasting infrastructure through what became, in the 2000s, the Indian Institute of Tropical Meteorology’s supercomputer-based seasonal-prediction system), Brazil (CPTEC, the Centro de Previsão de Tempo e Estudos Climáticos, which was significantly influenced by COLA visiting-scientist exchanges), South Korea, and Italy (where COLA collaborations contributed to the founding of the ICTP Earth System Physics group).19

Shukla’s training of doctoral students from developing countries was an explicit part of COLA’s mandate. He served as Ph.D. thesis adviser for “more than 20 students at MIT, the University of Maryland, and George Mason University,” with a deliberate weighting toward students from India, Brazil, and other developing countries who would return to their home institutions to build national climate-research capacity. The argument here was again partly architectural: training scientists on commodity Linux clusters meant that they could replicate the computational environment in their home countries when they returned, which they could not have done if they had been trained on $20-million Crays.

Shukla served as a Lead Author for the 2007 IPCC Fourth Assessment Report, sharing in the Nobel Peace Prize 2007 awarded jointly to the IPCC and Al Gore.20

Project Athena (2009-2010) and the late-IGES period

In November 2009 the National Science Foundation awarded COLA – with international collaborators ECMWF, the University of Tokyo, and Japan’s JAMSTEC – dedicated access to the Athena supercomputer, a 166-teraflop Cray XT4 system at the National Institute for Computational Sciences (NICS) at Oak Ridge, Tennessee, for a continuous six-month period beginning 1 October 2009. The principal investigator was James Kinter, by then COLA’s Director. Athena was at the time described as “four times as powerful as the Earth Simulator” – the Japanese vector machine that had held the top spot on the Top500 list earlier in the 2000s.21

Project Athena’s scientific output included high-resolution atmospheric simulations at 7-kilometre grid spacing – finer than any operational weather model of the period – with explicit (rather than parameterised) cloud-system resolution. The output paper is Kinter, J. L. III, B. Cash, D. Achuthavarier, et al. (2013) “Revolutionizing Climate Modeling with Project Athena: A Multi-Institutional, International Collaboration,” Bulletin of the American Meteorological Society 94(2), 231-245.22

Project Athena is a turning point in COLA’s institutional architecture. For the first time since 1993, the institution was running its principal scientific work on a Cray. The architectural commitment of 1993 – workstation clusters in preference to supercomputers – had been a commitment to owning throughput, not to avoiding supercomputers per se; when NSF made dedicated time on the largest supercomputer in the United States available, COLA used it. But the underlying argument that the standard operational rhythm of climate-prediction research is best served by owned commodity clusters did not change; the Athena allocation was an exceptional opportunity, not a new operational model.

The 2010 advocacy paper Shukla, J., T. N. Palmer, R. Hagedorn, B. Hoskins, J. Kinter, J. Marotzke, M. Miller, and J. Slingo (2010) “Toward a New Generation of World Climate Research and Computing Facilities,” Bulletin of the American Meteorological Society 91(10), 1407-1412 – co-authored by COLA leadership and by ECMWF’s senior climate-modeling staff (Tim Palmer, who at the time of this writing was about to leave ECMWF for Oxford) – argued for a new generation of climate-research supercomputers at “20 petaflops in the near term, 200 petaflops within five years, and 1 exaflop by the end of the next decade.” This is COLA’s most explicit institutional statement that for the next generation of climate-prediction work the workstation-cluster paradigm of 1993 was no longer enough.23

The 2013-2015 institutional move to George Mason

In May 2013 George Mason University announced that IGES and COLA would be “joining the University” and “located within” the College of Science. The dissolution of IGES as an independent corporation was decided in 2013; the COLA staff transitioned to GMU employment in stages from 2013 to 2015. In 2015 IGES/COLA staff were formally hired as GMU employees, and IGES as a corporation was dissolved.24 (The original prompt’s “2003” date for the GMU move is wrong; the correct date for the institutional consolidation is 2013-2015. Shukla had held a faculty appointment at GMU since 1994 and was Distinguished University Professor and the founding chairman of GMU’s Department of Atmospheric, Oceanic and Earth Sciences from the early 2000s, but COLA itself remained an IGES unit until 2014.)

The reason for the 2013 move appears to have been a combination of factors: the dissolution of the operational logic for an independent nonprofit research institute as the federal block-grant funding model became increasingly difficult to renew at the levels of earlier decades; GMU’s growing strength as a research university and its existing partnership with COLA through Shukla’s faculty appointment and the climate-dynamics Ph.D. programme; and the practical logic of consolidating a small institution’s administrative overhead with a much larger university’s. The move was contemporaneous with significant federal-funding pressure on climate-research nonprofits, which became amplified in the 2015-2016 controversy around Shukla’s signature on a letter to President Obama urging RICO investigation of fossil-fuel companies’ climate misinformation – a controversy that, while it post-dated the GMU consolidation, illustrated the political vulnerabilities of the independent-nonprofit model.25

Continuity of leadership was maintained throughout. Shukla served as the founding chairman of the GMU Department of Atmospheric, Oceanic and Earth Sciences (which he had founded in the early 2000s), and as Distinguished University Professor; James Kinter became (and remains, as of 2026) the Director of COLA at GMU and chair of the AOES department. The institutional identity – “Center for Ocean-Land-Atmosphere Studies” – was preserved unchanged; only the corporate parent shifted from IGES to George Mason.

The critical comparison with ECMWF

The contrast with the institution covered in Post 34 – the European Centre for Medium-Range Weather Forecasts – is the architectural inversion that defines this story. Both institutions were founded for the same broad scientific purpose: to do operational atmospheric prediction at a time horizon (medium-range weather for ECMWF, seasonal climate for COLA) that the standard national meteorological services could not adequately address. Both were founded by pooling resources that no individual investigator could command. Both have been institutionally durable for decades.

But the architectural choices were opposite. ECMWF, founded 1973 and operationally active from 1979, has run its operational forecasting on big-iron supercomputers from the start: Cray-1A serial number 9 (1978-1984), Cray X-MP/22 (1984-1985), Cray X-MP/48 (1985-1990), Cray Y-MP (1990-1994), Cray C90 (1994-1997), Fujitsu VPP700 (1997-2000), Fujitsu VPP5000 (2000-2003), IBM Cluster 1600 (2003-2009), IBM Power7 (2009-2014), Cray XC30 / XC40 (2014-2020), Atos BullSequana XH2000 (2020-). The institutional logic of ECMWF – pool sixteen national budgets, build one centralised facility, run twice-daily operational forecasts on a deadline that cannot slip – requires the kind of guaranteed-throughput, deadline-bounded computing that big-iron supercomputers were built to deliver. ECMWF’s operational rhythm (a ten-day forecast must complete in approximately one hour of wall-clock time so it can be disseminated by the deadline) is unforgiving in a way that climate-research workloads are not.26

COLA went the other way. Founded as a research institute, with no operational deadline, with a workload composed of large ensembles of long-running but throughput-tolerant integrations, with a budget that was generous by university-research-grant standards but small by national-supercomputing-centre standards, it bet that owned commodity hardware would serve its science better than rented Cray time. The bet paid off: through the IGES years (1993-2014), COLA produced a body of scientific work – the Predictability in the Midst of Chaos paradigm, the seasonal-prediction methodology, the COLA atmospheric and coupled models, the Athena project, the international-collaboration network – that was as influential in seasonal-to-interannual climate prediction as ECMWF’s was in medium-range weather forecasting. And it did so on hardware that, in aggregate, cost a small fraction of any one of ECMWF’s successive Crays.

The architectural divergence is, in retrospect, a permanent feature of the climate-research community. Operational prediction at deadline (ECMWF’s mission, the national weather services’ mission) requires guaranteed-throughput supercomputing. Research prediction in the ensemble-experiment style (COLA’s mission, the IPCC’s modelling-centre mission) is well served by owned commodity clusters. The two architectures are not in competition; they serve different missions, and both are still in active use in 2026.

What COLA did

The Center for Ocean-Land-Atmosphere Studies, in its independent existence from 1984 (Maryland) through 1993 (IGES) through 2014 (the GMU consolidation), demonstrated three things. First, that the seasonal-mean state of the tropical climate is predictable in a way that day-to-day weather is not – the Charney-Shukla hypothesis made operationally testable. Second, that a small dedicated nonprofit research institute could, on a budget of a few million dollars per year and on commodity Linux clusters of a few hundred thousand dollars each, do climate-prediction research at the scientific frontier. Third, that the Beowulf-pattern architectural revolution of 1994-2000 was the right computational substrate for ensemble-experiment climate science – not in opposition to but in parallel to the big-iron supercomputers that operational forecasting centres still use.

In 2026 COLA continues at George Mason University under James Kinter’s directorship, with approximately twenty-five staff, integrated GMU computing infrastructure, GPU-accelerated CESM-derived models, and an ongoing international-collaboration programme. Shukla, at age 82, is Distinguished University Professor Emeritus; his memoir A Billion Butterflies appeared in 2024.27

The architectural choice he made in 1993-1995 – workstation clusters in preference to a Cray – has, by 2026, become so universal in research-grade scientific computing that it is no longer recognised as a choice. Every modern climate-research cluster is descended from the Becker-Sterling Beowulf pattern; every Top500 supercomputer since 2017 runs Linux on x86 commodity hardware in some Beowulf-derived configuration. The architectural revolution is complete. COLA was one of the first institutions in atmospheric science to bet on it.

Footnotes

Key sources

  • Wikipedia “Jagadish Shukla” (en.wikipedia.org/wiki/Jagadish_Shukla)
  • Wikipedia “Beowulf cluster” (en.wikipedia.org/wiki/Beowulf_cluster)
  • COLA at George Mason University (cola.gmu.edu)
  • Sterling, T. L., et al. (1995) “BEOWULF: A Parallel Workstation for Scientific Computation,” ICPP 1995
  • Shukla, J. (1981) “Dynamical Predictability of Monthly Means,” J. Atmos. Sci. 38, 2547-2572
  • Shukla, J. (1998) “Predictability in the Midst of Chaos,” Science 282, 728-731
  • Shukla, J., et al. (2010) “Toward a New Generation of World Climate Research and Computing Facilities,” BAMS 91, 1407-1412
  • Kinter, J. L. III, et al. (2013) “Revolutionizing Climate Modeling with Project Athena,” BAMS 94, 231-245
  • Charney, J. G., and J. Shukla (1981) “Predictability of monsoons,” in Monsoon Dynamics
  • NASA Spinoff (2020) “Beowulf Clusters Make Supercomputing Accessible” (spinoff.nasa.gov/Spinoff2020/it_1.html)
  • GMU news (May 2013): COLA joins GMU
  • Shukla, J. (2024) A Billion Butterflies: A Life in Climate and Chaos Theory (memoir)

Cross-references in the post series

  1. Shukla, J. (2024) A Billion Butterflies: A Life in Climate and Chaos Theory, St. Martin’s Press / Macmillan, ISBN 9781250289209. The “primary school under a banyan tree” detail appears in multiple sources including American Kahani (2024) and the Star Tribune memoir review. ↩

  2. Shukla biographical: Wikipedia “Jagadish Shukla” (https://en.wikipedia.org/wiki/Jagadish_Shukla); GMU College of Science directory; American Kahani (2024) “Meet Meteorologist Jagadish Shukla, a Product of the Intellectual Pipeline That Makes America Great” (https://americankahani.com/lead-stories/meet-meteorologist-jagadish-shukla-a-product-of-the-intellectual-pipeline-that-makes-america-great/). Charney as Ph.D. supervisor at MIT 1976. ↩

  3. Shukla, J. (1981) “Dynamical Predictability of Monthly Means,” Journal of the Atmospheric Sciences 38(12), 2547-2572 (https://journals.ametsoc.org/view/journals/atsc/38/12/1520-0469_1981_038_2547_dpomm_2_0_co_2.xml). Affiliation: Goddard Laboratory for Atmospheric Sciences, NASA/GSFC, Greenbelt, MD. ↩

  4. Charney, J. G., and J. Shukla (1981) “Predictability of monsoons,” in Monsoon Dynamics, J. Lighthill and R. P. Pearce, eds., Cambridge University Press, pp. 99-108. The Charney-Shukla hypothesis is summarised in numerous subsequent reviews including Goswami et al. (2006) and Krishnamurthy and Shukla (2007). ↩

  5. COLA founding at Maryland 1984: GMU “Mason Trailblazer: Jagadish Shukla” (April 2022, https://www.gmu.edu/news/2022-04/mason-trailblazer-jagadish-shukla); Climate Audit “Shukla’s Gold” (2015) for Maryland-era staffing details; Wikipedia “Jagadish Shukla”. The original name “Center for Ocean-Land-Atmosphere Interactions” appears in early Maryland-era documents; “Studies” replaced “Interactions” by approximately 1988-1990. ↩

  6. Kinter biographical: COLA biography page (cola.gmu.edu/kinter/); ResearchGate profile. Princeton Ph.D. 1984 in geophysical fluid dynamics; NRC postdoctoral position at NASA Goddard; faculty position at University of Maryland; transition to IGES/COLA 1993; Director of COLA from approximately 2003 onwards. ↩

  7. IGES founding: Climate Audit (2015) “Shukla’s Gold” reports the corporation was incorporated in Maryland and the first NSF grants of $1.7 million were dated 6 January 1993. The “garage” anecdote appears in Wikipedia and in subsequent biographical pieces. ↩

  8. IGES block grants: Climate Audit (2015) “Shukla’s Gold” provides the financial details – 1994-1998 first block grant $2.25 million per year, NSF/NOAA/NASA joint funding, subsequent renewals 1999, 2004, 2009. The total $75-million figure for IGES grants over its existence is from the same source. Independent corroboration in Watts Up With That (2015) and the Daily Signal (2016). ↩

  9. Cray pricing in 1993: Cray Y-MP/8 list price approximately $25 million; Cray C90 approximately $35 million. NCAR allocation conventions documented in various NCAR annual reports of the period. ↩

  10. Beowulf-cluster economics: Sterling et al. (1995) ICPP paper for first-generation pricing; NASA Spinoff (2020) for the “10x cheaper than Cray” claim (https://spinoff.nasa.gov/Spinoff2020/it_1.html); various secondary sources for mid-1990s commodity-cluster pricing. ↩

  11. Beowulf 1 specifications: Sterling et al. (1995) “BEOWULF: A Parallel Workstation for Scientific Computation,” ICPP 1995, vol. I, pp. 11-14 (https://webhome.phy.duke.edu/~rgb/brahma/Resources/beowulf/papers/ICPP95/icpp95.html). Hardware: 16 Intel 80486 DX4 processors at 100 MHz, 16 MB DRAM per node, 500 MB hard drive per node, two channel-bonded 10baseT and 10base2 Ethernet networks. Performance: 60 MFLOPS sustained on the Prometheus CFD code. ↩

  12. Sterling et al. (1995), op. cit., direct quotes from the paper text. The CFD code Prometheus was a NASA Goddard atmospheric-science application; the comparison to the Cray T3D is in the same paper. ↩

  13. Sterling on resistance to Beowulf: NASA Spinoff (2020) (https://spinoff.nasa.gov/Spinoff2020/it_1.html). NASA’s Beowulf was inducted into the Space Technology Hall of Fame in 2022 (NASA NCCS news release, https://www.nccs.nasa.gov/news-events/nccs-highlights/beowulf). ↩

  14. The “Oh hell, just call it Beowulf” anecdote is in Sterling’s own retellings as captured in the NASA Spinoff (2020) feature and elsewhere. ↩

  15. Beowulf adoption in atmospheric science: NASA Spinoff (2020); Beowulf.org “Projects” page (https://www.beowulf.org/overview/projects.html); Bader (2025) “Linux and High-Performance Computing” (https://davidbader.net/publication/2025-b/2025-b.pdf). ↩

  16. Shukla, J. (1998) “Predictability in the Midst of Chaos: A Scientific Basis for Climate Forecasting,” Science 282(5389), 728-731 (https://www.science.org/doi/10.1126/science.282.5389.728). Affiliation: Center for Ocean-Land-Atmosphere Studies, Institute of Global Environment and Society, Calverton, MD. ↩

  17. Shukla, J., et al. (2000) “Dynamical Seasonal Prediction,” BAMS 81(11), 2593-2606. Note that Shukla also published a parallel companion paper in Quarterly Journal of the Royal Meteorological Society: Shukla, J., et al. (2000) “Dynamical seasonal predictions with the COLA atmospheric model,” QJRMS 126(569), 2265-2291 (https://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj.49712656714). ↩

  18. COLA coupled model: Kirtman, B. P., Y. Fan, and E. K. Schneider (2002) J. Climate 15(17), 2301-2320; Kirtman, B. P. (2003) “The COLA Anomaly Coupled Model: Ensemble ENSO Prediction,” Mon. Wea. Rev. 131(10), 2324-2341 (https://journals.ametsoc.org/view/journals/mwre/131/10/1520-0493_2003_131_2324_tcacme_2.0.co_2.xml). ↩

  19. International collaborations: GMU “Mason Trailblazer” (2022); COLA About page (https://cola.gmu.edu/aboutcola.php); Shukla’s IMO Prize 2007 citation. ↩

  20. IPCC AR4 contribution and 2007 Nobel Peace Prize: confirmed in multiple sources including GMU support-science feature, Wikipedia, and the IPCC’s published author lists. ↩

  21. Project Athena announcement: ORNL news release “NSF Dedicates Athena Supercomputer to Climate Research” (November 2009, https://www.ornl.gov/news/nsf-dedicates-athena-supercomputer-climate-research); NICS news release. Athena was a Cray XT4, 166 teraflops, 18,048 cores. The dedication was for six months starting 1 October 2009. ↩

  22. Kinter, J. L. III, et al. (2013) “Revolutionizing Climate Modeling with Project Athena,” BAMS 94(2), 231-245. The ECMWF participants included Tim Palmer (covered in Post 34 on the ECMWF Cray-1). ↩

  23. Shukla, J., T. N. Palmer, R. Hagedorn, B. Hoskins, J. Kinter, J. Marotzke, M. Miller, and J. Slingo (2010) “Toward a New Generation of World Climate Research and Computing Facilities,” BAMS 91(10), 1407-1412. Co-authorship with Palmer notable as Palmer was the architect of ECMWF’s Ensemble Prediction System (covered in Post 34). ↩

  24. GMU consolidation: GMU news (10 May 2013); Climate Audit (2015) for the dissolution timeline; multiple secondary sources for the 2014-2015 staff transition. ↩

  25. The 2015 RICO20 letter controversy: documented in multiple contemporaneous sources including the Daily Signal (2016, https://www.dailysignal.com/2016/03/02/house-probe-reveals-audit-detailing-climate-change-researchers-double-dipping-with-taxpayer-funds/) and Watts Up With That (2015). Post-dates the institutional consolidation. ↩

  26. ECMWF computing history: ECMWF supercomputer facility page (https://www.ecmwf.int/en/computing/our-facilities/supercomputer-facility); covered in detail in Post 34. ↩

  27. COLA in 2026: GMU College of Science directory; cola.gmu.edu; Shukla’s emeritus status; Kinter’s directorship. ↩