A ScyllaDB
Community
xCapture v3:
Efficient, Always-On
Thread Level
Observability with eBPF
Tanel Põder
Performance Nerd
Tanel Põder
A long-time computer performance nerd
■I made biolatency eBPF code 59x more efficient
■... when tested on 384 CPU server at 37M IOPS ☺
■P99CONF just keeps getting better, I’ll do my part!
■When not working, I hang out with 3..4 of our cats
Extended Thread State Sampling
Tracking + Sampling (not tracing)!
Two decoupled layers:
■ eBPF populating & maintaining a thread state
array
■ Keeps only the latest state change for each
thread
■ Sampling program independent from
population
■ Python/BCC, C, Rust/libbpf, eBPF iterators, etc...
■ Passive sampling without any probes or
xtop demo
dimensional performance analysis
xcapture architecture
xcapture
modular
architecture
Where to get it?
Open source, multiple performance tools:
■ https://0x.tools
■ https://github.com/tanelpoder/0xtools
Thank you! Let’s connect.
Tanel Põder
tanel@tanelpoder.com
@tanelpoder
tanelpoder.com

xCapture v3: Efficient, Always-On Thread Level Observability with eBPF by Tanel Poder