The high-performance computing (HPC) community has experienced a significant shake-up in its storage rankings following a crucial administrative decision by the IO500 Committee. Storage subsystems powered by Sugon’s ParaStor F9000 all-flash storage systems have been officially removed from the prestigious Production IO500 list. The demotion comes after independent scrutiny revealed that the submissions failed to satisfy stringent reproducibility requirements, which mandate comprehensive architectural documentation and general commercial availability. Consequently, Intel’s DAOS system, deployed on the powerful Aurora supercomputer at the Argonne National Laboratory, has reclaimed its position at the pinnacle of the production storage rankings.
The IO500 benchmark serves as the storage industry’s equivalent to the venerable TOP500 list. While TOP500 evaluates supercomputers strictly on their raw computational throughput measured in floating-point operations per second (FLOPS), the IO500 evaluates High-Performance Computing storage infrastructure based on Input/Output (I/O) performance, aggregate bandwidth, and metadata operations. As artificial intelligence workloads, massive data-driven scientific simulations, and exascale computing initiatives demand unprecedented data movement speeds, the IO500 list has emerged as the definitive benchmark for enterprise and research storage capabilities.
Background and Chronology of the ISC 2026 Submission Controversy
The controversy unfolded following the International Supercomputing Conference (ISC 2026), where SCNet submitted two distinct storage configurations leveraging Sugon’s proprietary ParaStor storage software and F9000 all-flash hardware architectures. These submissions immediately disrupted the status quo by posting staggering benchmark metrics that dramatically overshadowed existing benchmarks.
The larger of the two configurations, designated as SCNet AICS-A, executed the IO500 benchmark utilizing an expansive footprint of 500 client nodes equipped with a staggering 64,000 client processors. The system delivered a landmark IO500 score of 79,110.05, boasting an aggregate bandwidth of 26,888.39 GiB/s and a metadata performance rate of 232,754.76 kIOPS (thousand input/output operations per second).
Similarly, the smaller configuration, AICS-B, was submitted under the 10-client category utilizing 2,560 client processors. This setup posted an impressive score of 7,839.30, achieving 2,551.40 GiB/s of bandwidth and 24,086.69 kIOPS.
To put these figures into perspective, the SCNet AICS-A performance metrics substantially exceeded the benchmarks set by Argonne National Laboratory’s Aurora supercomputer. Aurora relies on a custom storage subsystem combining Intel Optane Persistent Memory modules, high-speed solid-state drives, and the open-source DAOS (Distributed Asynchronous Object Storage) file system. A comparable production benchmark for Aurora registered an overall score of 32,165.90, with 10,066.09 GiB/s of bandwidth and 102,785.41 kIOPS. The Sugon-powered AICS-A submission was roughly 2.46 times higher in overall performance than Aurora. In the 10-client category, the AICS-B submission outperformed Aurora by an approximate factor of 2.72.
Upon their initial acceptance, these impressive figures propelled the SCNet submissions directly to the number-one spots on both the main Production and 10-Client Production leaderboards, unseating Intel’s DAOS architecture, which had held the crown. However, the unprecedented nature of the scores quickly drew the attention of HPC researchers and benchmarking watchdogs, notably storage expert Glenn K. Lockwood, who publicly questioned the reproducibility and transparency of the underlying architecture.
The IO500 Committee Intervenes
Prompted by growing community scrutiny and an internal review of the submission criteria, the IO500 Committee took decisive action. The governing body formally transferred the Sugon ISC26 submissions from the Production List to the Research List, citing a failure to meet the highest standards of transparency.
In an official statement addressing the modification, the IO500 Committee explained: "After further review, the Sugon ISC26 submission has been transferred from the Production List to the Research List, as it did not satisfy the criteria for the highest level of Reproducibility due to the lack of widely available architectural details and limited general availability of the file system. Accordingly, the previous #1 position on the Production and Production 10-Client lists (Argonne’s DAOS system) has been restored."
The distinction between the Production list and the Research list is fundamental to the integrity of the IO500 benchmark. Much like the TOP500 list—where the top tiers frequently feature one-off, bespoke supercomputers constructed using experimental hardware—the IO500 permits specialized storage subsystems built upon exotic hardware and proprietary parallel file systems. Nevertheless, qualification for the Production list demands absolute adherence to rigorous reproducibility standards.
Defining the Standard: Reproducibility and Open Source vs. Proprietary Systems
The core controversy centers around what constitutes sufficient documentation and availability for a storage architecture to be deemed "reproducible."
For a system like Intel’s DAOS powering the Aurora supercomputer—which manages an immense 230-petabyte storage architecture—replication in a standard corporate data center or garage laboratory is practically unfeasible due to sheer scale and cost. However, DAOS satisfies the committee’s criteria because the underlying software is open source, freely downloadable, thoroughly documented, and accompanied by publicly accessible architectural blueprints, implementation guidelines, and explicit hardware and software specifications. Independent researchers can examine, audit, and theoretically deploy elements of the DAOS software stack on smaller scales.
In contrast, Sugon’s ParaStor architecture is strictly proprietary. According to the committee’s evaluation, the F9000 submission lacked the requisite public architectural disclosures, and the file system itself suffers from limited general availability. This lack of market transparency prevents independent verification and reproduction, rendering the system ineligible for the Production leaderboard under current guidelines.
Broader Implications for Proprietary HPC Architectures
This enforcement action highlights an ongoing structural challenge within international high-performance computing benchmarks regarding proprietary technologies, particularly those developed within regional technology ecosystems. Similar transparency challenges apply to other proprietary architectures, such as Huawei’s OceanFS and SuperFS file systems, which have traditionally been categorized under the Research IO500 list rather than the Production rankings.
For instance, the Research IO500 list is currently commanded by Pengcheng Laboratory’s CloudBrain system, which utilizes Huawei OceanStor A800 storage paired with the OceanFS file system. This setup achieved an astonishing IO500 score of 603,334.56, featuring 8,291.11 GiB/s of sequential throughput and an immense 43,903,983.64 KIOPS of random performance.
An interesting anomaly noted by technical observers within the leaderboard data is the discrepancy regarding classification flags. While submissions like Pengcheng Laboratory’s CloudBrain are explicitly marked as "proprietary" within the reproducibility status column of the Research IO500 list, the SCNet-A submission notably carried a "fully reproducible" badge despite its proprietary software foundations. The committee’s recent administrative correction addresses these classification inconsistencies, ensuring that the boundaries between verified, open-access production systems and closed, experimental research platforms remain strictly enforced.
Outlook for Future HPC Benchmarking
The demotion of the Sugon ParaStor F9000 systems underscores the growing importance of governance and data transparency in the era of exascale computing and artificial intelligence. As hardware capabilities scale into unprecedented territories, benchmarks like the IO500 must carefully balance the inclusion of cutting-edge industrial innovations with the foundational scientific necessity of reproducibility.
With Intel’s DAOS system successfully reinstated to the top of the Production IO500 and 10-Client Production lists, the HPC storage community remains focused on establishing clear, equitable standards that encourage technological advancement while preserving the credibility and verifiable nature of global supercomputing rankings.
