~/talk

Auto-instrumentation for GPU performance using eBPF

on YouTube 29m
Talkat DevConf.CZ 2025 · 2025-06-12..06-14· [machine-learning, observability]

speakers

name:Marc Tudurí
links:none on file
name:Dominik Süß
links:none on file

resources

video:https://youtu.be/gGe9QvSpSf8
slides:none on file

summary

Prototype uses eBPF probes on CUDA libcalls to capture kernel launches, host↔device memory copies and allocations at the CPU side, then exposes those events as Prometheus-exported kernel launch and memcpy metrics for Grafana. Implementation runs with minimal CPU overhead, is framework-agnostic for CUDA, open-source on GitHub, and shows per-container/kernel cardinality and memory-copy patterns. Limitations: no GPU-internal stats or per-kernel execution time and stack unwinding is partial when frame pointers are optimized away.

AI-written summary · about

chapters (13)

  1. 0:00GPU monitoring with eBPF
  2. 1:25Startup GPU struggles
  3. 3:59GPU failure risks
  4. 6:06Current tool limitations
  5. 9:58Grafana Beyla intro
  6. 10:41Zero-instrumentation tracking
  7. 13:39CUDA kernel costs
  8. 16:14GPU metrics in Prometheus
  9. 17:42GPU memory patterns
  10. 20:15GPU stack traces
  11. 21:38eBPF GPU limits
  12. 22:42GPU vs modern monitoring
  13. 24:02Q&A

related talks

~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.