Re: tcp proxy hackery

On Sat, Mar 08, 2008, Adrian Chadd wrote:
> Nothing terribly fancy; very, very ugly and not very well laid out; but
> cut #1 of a simple TCP proxy that I'm using for toying around with network
> IO benchmarking:
>
> http://www.creative.net.au/diffs/test1-320.tar.gz
>
> Edit Makefile.defaults to point to where libevent is hiding, and then 'make'.
> The destination host/port is hardcoded in tcpproxy-1/test.c ; I point it at a
> lighttpd/thttpd server and run apachebench against the proxy.

http://www.creative.net.au/diffs/test1-335.tar.gz

* The tcp proxy has been very naively threaded (defaults to two threads atm)
* I've included apachebench from apache-1.3, converted to use libevent
  and correctly complete the connect() sequence rather than hoping write-ready
  signified a successful connect. Its still single-threaded but it scales
  -enormously- better than before.

I'm able to push this to about 5000 req/sec, 8000 concurrent client connections
(so 16,000 concurrent TCP connections on the proxy) @ ~ 335mbit full-duplex on my
test setup. I'm not maxing out anything yet as my thttpd opteron server is
running at full steam.

Again, its all LAN only traffic; there's a 16k read-ahead buffer being used
here which settles the process at about 190 megabytes RSS. Thats not too bad
for what its doing, but obviously its going to be a lot more in the real world
if one goal is improved high-latency throughput.

Is anyone interested in helping me out by duplicating my setup and running
some basic tests? All you need is three sensible and recent servers, a gige
switch and some spare time. I'm still trying to debug the odd race condition
that showing up and resulting in mismatched content lengths.

Adrian

-- 
- Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support -
- $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -