Profiling and Benchmarking Zebra

Let's have a look at how we can inspect and evaluate Zebra's performance.

Profiling

To profile Zebra, you can use the samply profiler. Once you have it installed, you can run:

sudo samply record zebrad

where zebrad is the binary you want to inspect. You can then press Ctrl+c, and the profiler will instruct you to navigate your web browser to http://127.0.0.1:3000 where you can snoop around the call stack to see where Zebra loafs around the most.

Benchmarking

To benchmark Zebra consistently, you'll need to suppress unpredictable latency fluctuations coming from the network. You can do that by running two Zebra instances on your localhost: one that is synced up to the block height of your interest, and one that will connect only to the first instance.

To spin up the synced instance, you can run

cargo run --release -- --config /path/to/zebrad-synced.toml

with /path/to/zebrad-synced.toml pointing to the config below

# Config for a synced Zebra instance in a network-suppressed setup.

[network]
cache_dir = true
max_connections_per_ip = 1000
network = "Mainnet"

[state]
# cache_dir = "/path/to/.cache/zebra-synced"
delete_old_database = true
ephemeral = false

[sync]
checkpoint_verify_concurrency_limit = 1000
parallel_cpu_threads = 0

This config makes Zebra, among other things, accept quick reconnections from the same IP, which will be localhost. Without this setup, Zebra would quickly start treating localhost as a bad peer, and refuse subsequent reconnections, not knowing that they come from separate instances.

To spin up the second instance, first compile the version you want to benchmark:

cargo build --release

and run

time ./target/release/zebrad --config /path/to/zebrad-isolated.toml

with path/to/zebrad-isolated.toml pointing to the config below

# Config for an isolated Zebra instance in a network-suppressed setup.

[network]
listen_addr = "127.0.0.1:8234"
cache_dir = false
crawl_new_peer_interval = "10d"

initial_mainnet_peers = [
    "127.0.0.1:8233",
]

initial_testnet_peers = [
    "127.0.0.1:8233",
]

max_connections_per_ip = 1
network = "Mainnet"
peerset_initial_target_size = 1

[state]
# cache_dir = "/path/to/.cache/zebra-isolated"
delete_old_database = true
ephemeral = true
debug_stop_at_height = 10_000

This config makes Zebra:

1. connect only to the synced instance via localhost;
2. use an ephemeral state, so you can run the benchmark again;
3. stop at height 10_000.

Note that:

• You can adjust both configs to your liking.
• You can repeat the time command as many times as you need.
• You can use the two-instance setup for profiling as well.
• You will likely need to rebuild Zebra for each change you want to benchmark. To speed the build process up, you can link RocksDB dynamically, as described in the section on building Zebra.