Load Testing

[potatosalad]

I played around with some load testing last night using h2load and wanted to put the results up somewhere in case it's useful.

I also put together a few experiments with batched writing as well at potatosalad/hyper@chunked. The various patterns can be seen under lib/hyper/handlers.

The Hyper.Handlers.ChunkBatchSpawn implementation wound up having the most balanced performance behavior from preliminary testing, which is what is referred to as "hyperbeam chunked" below.

Changes made that increased performance slightly:

• Added batch_send_resp function
• Added Hyper.Batch process to enqueue and send writes to batch_send_resp
• Changed #[derive(NifMap)] to #[derive(NifTuple)] for the request struct
• Wrote a simplified proc_lib-based Hyper.Protocol to replace Hyper.Server

Note: There is something wrong with the HTTP/1.1 path on the rusterlium/hyper@batched branch that occasionally causes the server to slow down to the ~100 req/sec range. It's inconsistent, and seems to have a more pronounced effect as more pipelining is used (keeping the number of clients small, but streams high). Not sure what that's all about.

All tests were done on a 2017 MacBook Pro 13" with a 2-core 2.5GHz Intel Core i7 and 16 GB of LPDDR3 RAM.

Protocol	Clients	Streams	Server	Requests / sec
h1	100	10	hyper single_threaded	140,000
h1	10	100	hyper single_threaded	145,000
h2c	100	10	hyper single_threaded	100,000
h2c	10	100	hyper single_threaded	145,000
h1	100	10	hyper hello	220,000
h1	10	100	hyper hello	260,000
h2c	100	10	hyper hello	135,000
h2c	10	100	hyper hello	180,000
h1	100	10	hyperbeam master	65,000
h1	10	100	hyperbeam master	55,000
h2c	100	10	hyperbeam master	80,000
h2c	10	100	hyperbeam master	90,000
h1	100	10	hyperbeam batched	70,000
h1	10	100	hyperbeam batched	45,000
h2c	100	10	hyperbeam batched	95,000
h2c	10	100	hyperbeam batched	100,000
h1	100	10	hyperbeam chunked	80,000
h1	10	100	hyperbeam chunked	50,000
h2c	100	10	hyperbeam chunked	100,000
h2c	10	100	hyperbeam chunked	125,000
h1	100	10	cowboy handler	30,000
h1	10	100	cowboy handler	35,000
h2c	100	10	cowboy handler	17,000
h2c	10	100	cowboy handler	17,000
h2c	100	10	cowboy stream handler	35,000
h2c	10	100	cowboy stream handler	40,000
h1	100	10	go net/http	85,000
h1	10	100	go net/http	95,000
h2c	100	10	go net/http	45,000
h2c	10	100	go net/http	55,000
h1	100	10	h2o	155,000
h1	10	100	h2o	125,000
h2c	100	10	h2o	200,000
h2c	10	100	h2o	300,000
h1	100	10	nginx	95,000
h1	10	100	nginx	55,000
h2c	100	10	nginx	1,000
h2c	10	100	nginx	0

Server	Version	Command
hyper single_threaded	v0.12.24-12-g2114950c	cargo run --release --example single_threaded
hyper hello	v0.12.24-12-g2114950c	cargo run --release --example hello
hyperbeam master	09459c0	elixir -S mix run --no-halt
hyperbeam batched	97f2ad9	elixir -S mix run --no-halt
hyperbeam chunked	c7f9c48	elixir -S mix run --no-halt
cowboy handler	OTP/21.1,elixir/1.8.1,cowboy/2.6.1	elixir -S mix run --no-halt
cowboy stream handler	OTP/21.1,elixir/1.8.1,cowboy/2.6.1	elixir -S mix run --no-halt
go net/http	go1.11.4	go build server.go && ./server
h2o	2.3.0-DEV@330e23eb	h2o --conf "$(pwd)/h2o.conf"
nginx	nginx/1.15.9	nginx -c "$(pwd)/nginx.conf"

Client commands were:

# h1 (100 clients, 10 streams)
h2load --h1 --duration=5s --warm-up-time=1s --interval=1s --clients=100 --max-concurrent-streams=10 --requests=0 'http://127.0.0.1:3000/'
# h1 (10 clients, 100 streams)
h2load --h1 --duration=5s --warm-up-time=1s --interval=1s --clients=10 --max-concurrent-streams=100 --requests=0 'http://127.0.0.1:3000/'
# h2c (100 clients, 10 streams)
h2load --duration=5s --warm-up-time=1s --interval=1s --clients=100 --max-concurrent-streams=10 --requests=0 'http://127.0.0.1:3000/'
# h2c (10 clients, 100 streams)
h2load --duration=5s --warm-up-time=1s --interval=1s --clients=10 --max-concurrent-streams=100 --requests=0 'http://127.0.0.1:3000/'

Server implementations aren't fully documented anywhere (yet), but the cowboy and go related things are in potatosalad/ssm-stress-test.