RAM: transient peak ~22 GB during login/teleport (steady state 9 GB)

sky_line commented

2026-10-03 04:18:36 +00:00

Owner

Baltimare party session (83 avatars, 4K, 5090), ~9 min uptime after login plus two teleports:

VmRSS 9.0 GB (RssAnon 8.3 GB), VmHWM 22.3 GB. A transient spike about 13 GB above steady state, freed afterwards.
On a 16 GB machine this would swap or get OOM-killed even though steady state fits.

Suspects:

The decode burst at login/TP arrival: J2C decodes at full size before scale-down ("scaled-down 25202"; TextureFetchFullData fetches whole assets up to 128 MB). Many concurrent decode threads, each holding raw + decoded buffers.
Mesh decode/unzip bursts. The raw-mesh LRU is 256 MB, but the decode table and in-flight copies are outside that budget.
Staging / CPU texture copies before residency trims them.
Object cache / LLSD parsing of large region caches on arrival.

Plan:

Log RSS (statm) every 250 ms through login and a teleport; add per-subsystem counters for in-flight decode bytes (texture, mesh), staging and CPU tex copies, and correlate.
Cap concurrent in-flight decode bytes (a semaphore by bytes, not count).
Decode at the needed discard level instead of full size then downscale.
Throttle the arrival burst.

Target: peak RSS within ~1.5x of steady state.

Baltimare party session (83 avatars, 4K, 5090), ~9 min uptime after login plus two teleports: - VmRSS 9.0 GB (RssAnon 8.3 GB), **VmHWM 22.3 GB**. A transient spike about 13 GB above steady state, freed afterwards. - On a 16 GB machine this would swap or get OOM-killed even though steady state fits. Suspects: - The decode burst at login/TP arrival: J2C decodes at full size before scale-down ("scaled-down 25202"; TextureFetchFullData fetches whole assets up to 128 MB). Many concurrent decode threads, each holding raw + decoded buffers. - Mesh decode/unzip bursts. The raw-mesh LRU is 256 MB, but the decode table and in-flight copies are outside that budget. - Staging / CPU texture copies before residency trims them. - Object cache / LLSD parsing of large region caches on arrival. Plan: 1. Log RSS (statm) every 250 ms through login and a teleport; add per-subsystem counters for in-flight decode bytes (texture, mesh), staging and CPU tex copies, and correlate. 2. Cap concurrent in-flight decode bytes (a semaphore by bytes, not count). 3. Decode at the needed discard level instead of full size then downscale. 4. Throttle the arrival burst. Target: peak RSS within ~1.5x of steady state.

sky_line added the

perf

bug

labels

2026-10-03 04:18:36 +00:00

sky_line referenced this issue

2026-10-03 23:47:08 +00:00

GPU-ready texture cache: store block-compressed textures on disk (skip J2C decode + encode on reload) #274