Do C10K testing with gevent | Hacker News Reader

Do C10K testing with gevent

engineering.monetate.com

30 pointsphillytom14 comments

C1M is the new C10K.

C10K is not even close to being a real problem any more. I expect my phone can handle 10K connections.

Making a million outbound connections, on the other hand, requires a little more care. Mostly because you need 16+ IPs. I ended up writing a small tool in C with libevent to make 1M http connections: http://www.metabrew.com/article/a-million-user-comet-applica...

you're only going to need 16+ IPs if you're making those outbound connections to the same destination IP & port... which seems unlikely.

You can do TCP listen on 0.0.0.0:80 (any IP on port 80), but I never saw TCP listen on XXX.XXX.XXX.XXX:* (specific IP with any port) ;)

1. I already solved C1M problem, so C10K and C100K looks like a non-problem to me.

EDIT: metabrew has very good tutorial - which should be a starting point for everybody new to this space!

2. The problem is, that Amazon EC2 doesn't support multiple IP aliases per instance. To generate 1M connections, I had to launch a cluster of 17 instances to generate 1M clients. On physical server with a lot of RAM, you can just create 17 IP aliases.

3. Tsung is a great tool - but hard to control and has it's own limits. So if you need to do large scale loadtesting - better write your own tool.

I have a stupid question here. When you have a large memory machine, why can't you serve 1M clients on 1 ip without IP aliases?

I asked this question because if one uses 17 instances to handle 1M connections, for each instance is around 50k. I am more interested in knowing the possibility and performance for one machine to handle 1M connections.

Port numbers are 2byte values. To track a connection as an IP:port number combination(the way your typical OS does it) you need multiple IPs to give you enough port numbers.

each socket can be uniquely identified by 4-tuple: {LocalIP,LocalPort,ServerIP,ServerPort}

LocalIP:LocalPort -> ServerIP,ServerPort

So even if number of ports is limited by 64K, total number of incoming sockets per IP address isn't.

Yes, but when you are load testing a specific server, as in this case, you are bound to a specific serverIp and serverPort.

And thus, the total amount of sockets you can have open to that server is limited by the amount of IP's your machine has.

If we are testing a web server on port 80, such that ServerIP,and port are fixed, then the only way you are getting more than 64K connections is through multiple client IP addresses.

In the real world your 1M connections will be coming from ~1M clients, so there's no problem. When doing load testing with one client and one server you run into port number limits; this is most realistically solved by using multiple IP addresses on the client (or multiple clients if your cloud is crippled).

the real problem is not to generate 1M outgoing connections, but to handle 1M incoming connections on a single node (C1M), while in addition also doing some meaningful processing.

It's a solved problem for me. That is why I don't understand why many high profile realtime web startups (like Convore), still using inferior solutions, like Python eventlib or gevent, which at most let them handle hundreds-to-thousands of clients.

EDIT: metabrew has very good tutorial - which should be a starting point for everybody new to this space!

Are you referring to http://www.metabrew.com/article/a-million-user-comet-applica...

http://www.metabrew.com/article/a-million-user-comet-applica...

By Richard Jones of the Last.fm?

Thank you for your replying. The original C10k from Dan Kegel was a guideline for me when I worked on project during 2003 - 2006. I am wondering what is the latest progress in this area. I am glad this problem is solved to you. I will check out niver tech in future when the problem comes up.

I actually was pretty fond of this:

https://github.com/newsapps/beeswithmachineguns

I don't remember how I came across it but it's the same concept: fire up a swarm of EC2 instances to pound the shit out of your site.

He doesn't seem to have tried Tsung which is a good solution for this scale of testing.

Reply on news.ycombinator.com