Industry

Sugon's ParaStor F9000 Demoted From IO500 Production Rankings Over Reproducibility Concerns

The IO500 Committee has relocated Chinese storage systems from its top-tier Production list to Research, citing insufficient documentation and limited availability. Intel's Aurora reclaims the Production crown.

3 min read
Sanctioned Chinese supercomputer maker stripped of IO500 benchmark crown, Intel-powered Aurora retakes the lead — record-breaking ParaStor F9000 storage system doesn't meet reproducibility requirements

Despite the demotion, Sugon's ParaStor F9000 remains among the world's most capable storage subsystems. The IO500 Committee determined that storage configurations built on Sugon's ParaStor F9000 all-flash platform failed to satisfy reproducibility standards required for Production-tier rankings, prompting their transfer to the Research IO500 list. Glenn K. Lockwood flagged the issue, which centers on inadequate disclosure of system architecture and restricted accessibility. Following this decision, Intel's Aurora regained the top position on the Production list.

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.

IO500 Committee

The IO500 benchmark functions as the storage-focused equivalent to TOP500, measuring HPC storage system performance through input/output metrics rather than computational throughput. It evaluates systems based on aggregate bandwidth and I/O operations.

At ISC 2026, SCNet presented two ParaStor-based configurations. The larger AICS-A system executed the IO500 test across 500 client nodes commanding 64,000 processors, posting an IO500 score of 79,110.05 alongside 26,888.39 GiB/s bandwidth and 232,754.76 kIOPS metadata throughput. The smaller AICS-B variant used 10 clients with 2,560 processors, achieving 7,839.30 points, 2,551.40 GiB/s, and 24,086.69 kIOPS. Both identified Sugon as the storage vendor and ParaStor as the file system.

These figures substantially outpaced Argonne National Laboratory's Aurora system, which employs a custom storage stack combining Intel Optane Persistent Memory modules, SSDs, and the DAOS file system. Aurora's comparable Production submission registered 32,165.90 points with 10,066.09 GiB/s bandwidth and 102,785.41 kIOPS—meaning AICS-A's score was approximately 2.46 times higher. In the 10-client category, AICS-B's 7,839.30 exceeded Aurora's 2,885.57 by roughly 2.72 times. Upon initial acceptance to Production lists, the SCNet submissions displaced Aurora from both the main Production and 10-Client Production top positions.

Like TOP500, the IO500 list accommodates entirely custom storage subsystems featuring specialized hardware and bespoke parallel file systems. However, Production-tier submissions must satisfy the highest reproducibility threshold: sufficient architectural documentation and general availability of the underlying file system to enable independent verification and replication of results.

While reproducing Aurora's 230PB storage subsystem independently remains impractical for most organizations, its design and software remain openly examinable and reproducible at smaller scales. DAOS operates as open-source software with extensive documentation, publicly disclosed architecture, implementation specifics, and transparent hardware and software prerequisites.

Sugon's ParaStor presents a contrasting profile: the proprietary system lacks transparent documentation. The IO500 Committee determined the F9000 submission provided insufficient publicly accessible architectural information and that the file system had constrained general availability, preventing the independent examination and reproduction necessary for Production-list qualification.

This reproducibility standard similarly affects other proprietary Chinese architectures, including Huawei's OceanFS and SuperFS, which appear on the Research list. The Research IO500 rankings are presently led by Pengcheng Laboratory's CloudBrain system running Huawei OceanStor A800 storage with OceanFS, which posted an IO500 score of 603,334.56, 8,291.11 GiB/s sequential throughput, and 43,903,983.64 KIOPS random performance.

Google Preferred Source

A notable inconsistency exists in the Research IO500 list's reproducibility designations. While Pengcheng Laboratory's CloudBrain and CloudBrain-II submissions carry 'proprietary' badges in the reproducibility column, the SCNet-A submission displays a 'fully reproducible' designation.

Source: Tom's Hardware

Source: Tom's Hardware · Reporting supplemented by The Silicon Ledger staff.