Benchmark Support - setting concurrency, num-requests, & time-limit.

[HelloGrayson]

Their are lots of times where I've found myself wanting to do a quick and dirty stress test from tcurl. Let's add Apache Bench-like functionality so that we can easily run concurrent traffic from the command line.

Starting with a --concurrency flag would go a long ways towards this goal.

# issue 100k requests at 1k concurrency
tcurl hyperbahn --health --concurrency=1000 --num-requests=100000

# issue requests for 1m at 1k concurrency
tcurl hyperbahn --health --concurrency=1000 --time-limit=60

[Raynos]

Please re-use either:

• https://github.com/uber/tchannel-node/blob/master/test/lib/batch-client.js
• https://github.com/uber/hyperbahn/blob/master/test/lib/time-series-cluster.js#L303-L682

Probably the hyperbahn one. It has a #lotoffeatures

[HelloGrayson]

Golang's sniper also has a nice interface: https://github.com/lubia/sniper

[Raynos]

I would love to have this feature btw; if someone can bang it out then mega :+1: should be pretty easy if you steal the BatchClient from hyperbahn.

[ShanniLi]

Do we want to also add a time interval between each burst?

[Raynos]

What hyperbahn batch client does is:

• N reqs in parallel;
• wait M ms;
• repeat.

I dont know how that measures with "concurrency"; you would probably be like "if concurrency 1000 then divide by 100; do batches of N=100reqs; M=100ms"

You probably also want a flag for "flushing/blocking" etc.

The BatchClient in hyperbahn does not wait; it will just slam you on an interval and not care about response or "maximum requests in flight".

[ShanniLi]

I think @breerly means: concurrency == N.

The reason I ask about the time interval is that I want some way to send requests at a certain QPS. This will also be useful for testing rate limiting. I have my own private integration test for that now. :)

[Raynos]

I think tests where we send Nk at the start are useless.

The dials you want is:

• req/s
• total reqs or total time
• maximum reqs in flight
• delay

tcurl --requests 100000 --rate 1000 #100k @ 1k req/s
tcurl --time 60 --rate 1000 #60s @ 1k req/s
tcurl --requests 100000 --rate 1000 --concurrency 1000 #100k @ 1k req/s bound concurrency to 1000
tcurl --requests 100000 --rate 1000 --delay 250 #100k @ 1k req/s with 250ms batching delay

Note: default concurrency of {rate} and default delay of 100ms

I may be completely wrong here.

[ShanniLi]

@breerly I think rate is what you want. Do we also need concurrency?

[Raynos]

we totally need concurrency :) but maybe not v0.

[ShanniLi]

@Raynos Sounds good, I will start with v0 without concurrency, i.e., set concurrency = rate.

[HelloGrayson]

Sounds good to me guys :)