(gpu-v2026): ASW domain sizes do not match paper-stated 32-64 nodes per ASW

[Varsius]

Regarding cluster-trace-gpu-v2026:

When reconstructing the cluster topology from server_hourly by counting unique server_id values per (asw_id, cluster_id) pair, I observed the following:

• Total unique servers: 38,048 (paper states 37,707)
• Unique cluster_id values: 17
• Unique (asw_id, cluster_id) pairs (ASW domains): 2,522
• Servers per ASW domain: min=1, max=136, median=6, avg=15.0

Section 2 of the "Heterogeneity at Hyperscale" paper states:

"Each ASW aggregates 32–64 nodes, so a single ASW domain spans roughly 256–512 GPUs."

(1) Could you please clarify whether asw_id corresponds to individual physical access switches, or a different grouping? Do you have a possible explanation why these numbers don't aligned with the 32-64 nodes per ASW?
(2) Could you please clarify whether cluster_id corresponds to the paper-stated aggregation switches (AGG)?

[mental2008]

Thanks for your attention. The discrepancy comes from comparing different scopes:

1. The 37,707-server number in the paper is the peak number of GPU servers observed concurrently in an hourly snapshot. The 38,048 value is the number of distinct server_id values seen over the entire six-month trace. Server additions, removals, and replacements during this period make the latter slightly larger than the hourly peak.
2. asw_id identifies an anonymized access-switch domain. The 32–64 nodes described in the paper refer to all physical nodes connected under an ASW, while asi_opensource_server_hourly contains only the GPU servers covered by the released trace. Therefore, the number of visible GPU servers under an ASW can be much smaller than 32. Conversely, counting distinct server_id values across all 185 days can produce a value larger than 64 because it combines fleet changes over time. A closer comparison would count servers separately for each (day, hour, cluster_id, asw_id) snapshot, although it would still cover only the GPU-server subset in the dataset.
3. cluster_id is an anonymized logical cluster identifier. It does not correspond to the AGG shown in the paper. AGG-level identifiers are not included in the released trace.

[Varsius]

This makes sense. Thanks for the clarifications.