Computer Science > Distributed, Parallel, and Cluster Computing
Title:Size-Biased Admission in Print-Gated GPU Clusters
Abstract:A print-gated scheduler issues a placement request only after one allocation sheet per requested GPU has reached the printer's output tray. If a jam invalidates the partly printed ticket, wide gangs need more attempts and rejoin the spool more often. We derive the offered load of this admission stage and evaluate it by discrete-event replay on a 64-GPU cluster model. At a 7% per-feed jam probability, eight-GPU tickets discard 2.47 ejected sheets per admission. Compared with an otherwise identical jam-free printer, jams reduce the scheduler's 95th-percentile placement wait from 110 to 93 s while increasing the client's 95th-percentile wait from 121 to 561 s. GPU utilization differs by 0.35 percentage points. A 15 min interruption leaves the placement queue empty and all GPUs idle while an average of 174 requests wait in the spool. Fitting two accelerator records per sheet retains the reprint rule and lowers the client tail to 135 s. The client clock includes the delay at the print spool.
Submission history
From: Mara Venn[v1] Fri, 20 Nov 2026 21:17:05 UTC (176 KB)