stacks

stacks is a bpf-based tool to collect stack traces together with resource usage. as of now it can collect data for the following events:

• cpu profiling
• scheduler events
• changes in rss
• block io and vfs events
• tcp/udp send/recv events

all events can be collected with user/kernel stacks and symbolized and generally doesn't require changes in the software. one required change might be to compile with force-frame-pointers=yes, and it is also beneficial to keep debug=1.

additionally, tracing-stacks is a basic extension for popular tracing crate. if enabled it will label all os events that happened within the span with correct span and span name. TODO add an example how it is useful (e.g in pprof or traceviewer)

the data is collected into parquet files. it can be later exported into pprof or chrome traceviewer with stacksexport tool or explored with jupyter.

i provided several examples how it can be used in explore directory.

Using

To compile you need the following dependencies:

sudo apt install -y build-essential autoconf clang-15 flex bison pkg-config autopoint
sudo ln -s /usr/include/asm-generic /usr/include/asm
sudo rm -f /bin/clang
sudo ln -s /usr/bin/clang-15 /bin/clang

And then install with cargo:

cargo install --path stacks
cargo install --path stacksexport

For example i will collect stacks from running stacks, code and firefox commands. The command below will be doing collecting user stacks at 99hz rate, user stacks every 29th rss growth event, kernel and user stack for cpu switch event and user/kernel stacks for write/read from block device.

sudo ./target/release-debug/stacks stacks code firefox -p profile:u:99,rss:u:29,switch:uk,block:uk

After collecting data it can be viewed with stacksexport, also make to install pprof.

profile

stacksexport pprof -b ./target/release-debug/stacks ./stacksexport/sql/pprof/cpu_ustacks_for_buildid.sql

By providing -b ./target/release-debug/stacks to stacksexport it is also possible to extend collected data with source code information for nicer UX.

rss

stacksexport pprof -b ./target/release-debug/stacks ./stacksexport/sql/pprof/rss_ustacks_growth_for_buildid.sql

One important detail about rss is that events will be captured when program request memory pages from os, so for example it will show rss growth when vector is initialized, but will show when element is written.

offcpu

stacksexport pprof -b ./target/release-debug/stacks ./stacksexport/sql/pprof/offcpu_stacks_for_buildid.sql

Offcpu events capture last stack traces before thread was switched off by scheduler, in this profile program waits for ring buffer events in epoll.

block

There are two different profiles for block operations, by amount and duration. And it is possible to generate different profiles for blk_read and blk_write (see queries).

stacksexport pprof -b ./target/debug/examples/writer ./stacksexport/sql/pprof/blk_ustack_duration_for_buildid.sql
stacksexport pprof -b ./target/release-debug/stacks ./stacksexport/sql/pprof/blk_ustack_amount_for_buildid.sql

This is not a particularly interesting example, capture from e2e/examples/writer with fsync after each write.

Tips

Splitting large profiles for faster pprof generation.

Query in the exporter can be updated to include timestamp:

where
    kind = 'rss'
    and buildid = decode('?buildid', 'hex')
    and timestamp < 1726573942947884522

Profiles are collected in relatively small batches, as at the time this is a unit of paralelizatin in datafusion. They can be registered individually:

stacksexport pprof -b ../go-spacemesh/build/go-spacemesh ./stacksexport/sql/pprof/rss_ustacks_growth_for_buildid.sql -r "/tmp/stacks/10/STACKS-1-*"