Topic: Mongrel testing with httperf

I've been experimenting with my new slice on Slicehost (best thing evs, OMG).

I've been testing the slice against my current host which only supports 1 mongrel instance (CPanel on Apache, so no mod_proxy).

Using httperf
rate = 50
num-conns = 2000


4 Mongrel Instances:

Connection rate: 49.7 conn/s (20.1 ms/conn, <=27 concurrent connections)
Request rate: 15.5 req/s (64.4 ms/req)
Reply rate [replies/s]: min 15.0 avg 15.5 max 18.2 stddev 1.1 (8 samples)

Single Mongrel Instance:
Connection rate: 49.9 conn/s (20.0 ms/conn, <=24 concurrent connections)
Request rate: 15.5 req/s (64.5 ms/req)
Reply rate [replies/s]: min 14.6 avg 15.5 max 19.6 stddev 1.7 (8 samples)

Can this be right?
Is it possibly the client that is the bottleneck here?

Last edited by tobyhede (2007-11-27 21:04:56)

Toby Hede
===================================================
FiniteStateMachine - Software Development for Social Networks
===================================================

Re: Mongrel testing with httperf

There are a few things that might be happening. I'm new to httperf too so YMMV, CYA, IANAL, etc. tongue

1. Client - hit another site (preferably one you own) with httperf and see what you get
2. Caching - if pages get cached the first hit will do the work and then it'll just serve from the cache. If pages are being served from the cache then apache will just serve them up and you'll be testing it instead of mongrel.
3. Actual Load - 4 mongrels rendering index.html is still just mongrel rendering index.html. How are you running httperf? Is it actually making mongrel do some significant work?

Re: Mongrel testing with httperf

I've been running the tests off one of my servers, instead of my laptop (running wireless connection) and the results are MUCH more consistent.

Toby Hede
===================================================
FiniteStateMachine - Software Development for Social Networks
===================================================