Skip to content

Latest commit

 

History

History
76 lines (54 loc) · 1.77 KB

performance.md

File metadata and controls

76 lines (54 loc) · 1.77 KB
Raw

Measuring throughput with benchmarking tool

  • start the broker
RUST_LOG=rumq_broker=info cargo run --release
  • run the tool with the configuration you like
./benchmark --help
./benchmark -c 10 -m 10000 (open 10 connection with 10000 publishes each)
  • results look like this
tekjar@thinkpad:~/Workspace/rumq/utils/benchmark$ ./benchmark -c 10 -m 50000
  99% |||||||||||||||||||||||||||||||||||||||.  [29s:0s] 

Id = bench-0 Total Messages = 50000 Average throughput = 1.6353620750123208 MB/s
Id = bench-1 Total Messages = 50001 Average throughput = 1.6310719156488749 MB/s
...

 Total Messages =  500005 Average throughput =  16.03808488093616 MB/s

Findings

  • There doesn't seem to be any improvement wrt cpu and throughput using BufStream
./benchmark -c 10 -m 10000 (open 10 connection with 10000 publishes each)
  • Using codec on top of Read has a huge perf benefit over async MqttRead/Write

There is 8X improvement in throughput

// https://docs.rs/tokio/0.2.6/src/tokio/io/util/buf_reader.rs.html#117
// TODO What happens when socket receives only 4K of data and there is no new data?
// TODO Also might have to periodically flush to send mqtt packets which are part of
// partially filled write bufers 
let mut stream = BufStream::new(&mut self.stream);

Moooree tools

#####Flamegraph

RUSTFLAGS="-g -C force-frame-pointers=y" cargo flamegraph rumq-cli

#####Perf

Build in release mode with debug symbols (-g) and frame pointers enabled

RUSTFLAGS="-g -C force-frame-pointers=y" cargo build --release
perf record -g <binary>
perf report -g graph,0.5,caller

References