cloud-hypervisor/docs/profiling.md

Profiling

perf can be used to profile the cloud-hypervisor binary but it is necessary to make some modifications to the build in order to produce a binary that gives useful results.

Building a suitable binary

Modify the Cargo.toml file to add debug = 1 to the [profile.release] block. It should look like this:

[profile.release]
lto = true
debug = 1

This adds the symbol information to the release binary but does not otherwise affect the performance.

The binary must also be built with frame pointers included so that the call graph can be captured by the profiler.

$ cargo clean && RUSTFLAGS='-C force-frame-pointers=y' cargo build --release

Profiling

perf may then be used in the usual manner:

e.g.

$ perf record -g target/release/cloud-hypervisor \
        --kernel ~/src/linux/vmlinux \
        --pmem file=~/workloads/focal.raw \
        --cpus boot=1 --memory size=1G \
        --cmdline "root=/dev/pmem0p1 console=ttyS0" \
        --serial tty --console off \
        --api-socket=/tmp/api1

For analysing the samples:

$ perf report -g

If profiling with a network device attached either the TAP device must be already created and configured or the profiling must be done as root so that the TAP device can be created.

Userspace only profiling with LBR

The use of LBR (Last Branch Record; available since Haswell) offers lower overhead if only userspace profiling is required. This lower overhead can allow a higher frequency of sampling. This also removes the requirement to compile with custom RUSTFLAGS however debug symbols should still be included:

e.g.

$ perf record --call-graph lbr --all-user --user-callchains -g target/release/cloud-hypervisor \
        --kernel ~/src/linux/vmlinux \
        --pmem file=~/workloads/focal.raw \
        --cpus boot=1 --memory size=1G \
        --cmdline "root=/dev/pmem0p1 console=ttyS0" \
        --serial tty --console off \
        --api-socket=/tmp/api1