From e02d964ff61461f5717e026156133dfaaa6a40c8 Mon Sep 17 00:00:00 2001 From: zainhas Date: Mon, 13 Jul 2026 17:57:04 +0000 Subject: [PATCH 1/2] sync together-gpu-clusters with mintlify-docs#1196 Update the Storage Performance active health check description to match the docs changes in togethercomputer/mintlify-docs#1196 (DX-785): per-tier fio thresholds and fio configurations, and the removal of the Skipped result state (shared-tier now self-provisions a temporary isolated volume, no shared-volume PVC precondition). --- skills/together-gpu-clusters/references/cluster-management.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/skills/together-gpu-clusters/references/cluster-management.md b/skills/together-gpu-clusters/references/cluster-management.md index f23cfe5..b2e943a 100644 --- a/skills/together-gpu-clusters/references/cluster-management.md +++ b/skills/together-gpu-clusters/references/cluster-management.md @@ -352,7 +352,7 @@ Nodes showing "Tests Failed" are not added to the cluster until repaired. - NVBandwidth: CPU-to-GPU, GPU-to-CPU bandwidth, GPU-CPU latency **Storage:** -- Storage Performance: `fio` sequential read/write throughput plus a checksummed write/read-back for data-integrity validation against shared and local volumes attached to the cluster. Results include a **Skipped** state (in addition to Passed/Failed) -- on Kubernetes clusters the shared-storage portion reports **Skipped** with message `Shared volume PVC not created` until a shared-volume PVC exists (see the [Storage](#storage) section). +- Storage Performance: `fio` sequential read/write throughput plus a checksummed write/read-back for data-integrity validation against the cluster's shared and local storage tiers. Thresholds are per-tier: shared expects ≥ 10 GiB/s read and ≥ 5 GiB/s write (1 MiB blocks across 64 jobs at iodepth 32); local expects ≥ 2 GiB/s read and ≥ 1 GiB/s write (1 MiB blocks across 16 jobs at iodepth 16). The shared-tier check self-provisions a temporary isolated volume, so it no longer requires a pre-existing shared-volume PVC and no longer reports a **Skipped** result. ### Node Repair From 8160ee8b9fa9de1871c5352166330cd22841d9e2 Mon Sep 17 00:00:00 2001 From: Neeraj Shah Date: Mon, 13 Jul 2026 11:08:41 -0700 Subject: [PATCH 2/2] Trim storage-performance blurb: drop FIO thresholds and Skipped tombstone Remove the per-tier FIO throughput thresholds and test configuration, plus the tombstone note about the shared-tier check self-provisioning a volume. Co-Authored-By: Claude Opus 4.8 (1M context) --- skills/together-gpu-clusters/references/cluster-management.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/skills/together-gpu-clusters/references/cluster-management.md b/skills/together-gpu-clusters/references/cluster-management.md index b2e943a..9ec8393 100644 --- a/skills/together-gpu-clusters/references/cluster-management.md +++ b/skills/together-gpu-clusters/references/cluster-management.md @@ -352,7 +352,7 @@ Nodes showing "Tests Failed" are not added to the cluster until repaired. - NVBandwidth: CPU-to-GPU, GPU-to-CPU bandwidth, GPU-CPU latency **Storage:** -- Storage Performance: `fio` sequential read/write throughput plus a checksummed write/read-back for data-integrity validation against the cluster's shared and local storage tiers. Thresholds are per-tier: shared expects ≥ 10 GiB/s read and ≥ 5 GiB/s write (1 MiB blocks across 64 jobs at iodepth 32); local expects ≥ 2 GiB/s read and ≥ 1 GiB/s write (1 MiB blocks across 16 jobs at iodepth 16). The shared-tier check self-provisions a temporary isolated volume, so it no longer requires a pre-existing shared-volume PVC and no longer reports a **Skipped** result. +- Storage Performance: `fio` sequential read/write throughput plus a checksummed write/read-back for data-integrity validation against the cluster's shared and local storage tiers. ### Node Repair