Re: performance

"Steven Tower" <[email protected]> Mon, 20 May 2002 11:36:45 -0400
Newsgroups gmane.linux.ngpt.user
Message-ID <038a01c20014$28cd6310$1400000a@towerpc>
Bill,

So far my tests with siege against apache2 yield the worse results then with
siege against tux.

Siege compiled with ngpt eats signifigantely more cpu time then with
linuxthreads and siege (ngpt) against apache2 compiled with ngpt yields the
worst results of all.

What seems to be happening is that siege when using ngpt is starving the app
I am hitting.

So for example tux sieged with 60 concurrent threads delivers 2200 ops per
second with a concurrency of 59 connections with Linuxthreads, with ngpt it
averages 330 ops per second with a concurrency of 30-59 (it varies that
widely) connections.

Against Apache2 (ngpt) and siege (ngpt) I get 59 ops per second with a
concurrency of 10-20 connections.

Against Apache2 (ngpt) and siege (linuxthreads) I get 900 ops per second
with a concurrency of about 30.

Against Caucho Resin (Blackdown JDK/Resin 2.1.0) and siege (linuxthreads) I
get 1100 ops per second with a concurrency of about 55.

Against Caucho Resin (Blackdown JDK/Resin 2.1.0) and siege (ngpt) I get 300
ops per second with a concurrency of about 35-40.